[{"data":1,"prerenderedAt":843},["ShallowReactive",2],{"blog-post-kups-a-molecular-simulation-engine-for-the-ai-era":3},{"_id":4,"_updatedAt":5,"acknowledgements":6,"body":18,"credits":754,"excerpt":34,"hero":755,"publishedAt":769,"related":770,"seo":828,"slug":830,"thumbnail":56,"title":831,"headings":832,"footnotes":842,"hasMath":829},"post-medium-b213963a2359","2026-11-13T13:26:13Z",[7],{"_key":8,"_type":9,"children":10,"markDefs":16,"style":17},"6a3c0e8acf32","block",[11],{"_key":12,"_type":13,"marks":14,"text":15},"535a24304769","span",[],"With thanks to Elias Moubarak, Joe Morrow, Pim de Haan, Hannah Openshaw, Rhiannon Tait, Jessica Valdez, Ruth Barnett, Max Welling, CuspAI Team.",[],"normal",[19,28,37,59,67,90,101,116,133,143,151,159,199,218,233,242,253,272,291,297,306,321,329,337,351,366,374,388,397,405,418,429,438,453,466,474,511,533,542,557,565,584,593,604,619,628,646,651,684,692,701,726,734],{"_key":20,"_type":9,"children":21,"markDefs":27,"style":17},"ce499ceb250a",[22],{"_key":23,"_type":13,"marks":24,"text":26},"6e71d3b96804",[25],"em","Technical blog",[],{"_key":29,"_type":9,"children":30,"markDefs":35,"style":36},"c67d186c9466",[31],{"_key":32,"_type":13,"marks":33,"text":34},"14c813d3d9d4",[],"Molecular simulations are the backbone of drug discovery and materials science, but they’re notoriously complex.",[],"lead",{"_key":38,"_type":9,"children":39,"markDefs":53,"style":17},"aa58cba47833",[40,44,49],{"_key":41,"_type":13,"marks":42,"text":43},"ac1a8b27017b",[],"For example, the pharmaceutical industry ",{"_key":45,"_type":13,"marks":46,"text":48},"a4e3eb0a9959",[47],"56409e20149c","spends 15% of its revenue on R&D",{"_key":50,"_type":13,"marks":51,"text":52},"d48b6825f445",[]," — more than 20% of the total R&D spend across all other sectors combined — and molecular dynamics is one of the heaviest computational line items in that budget.",[54],{"_key":47,"_type":55,"internalSlug":56,"type":57,"url":58},"link",null,"external","https:\u002F\u002Fwww.mckinsey.com\u002Findustries\u002Flife-sciences\u002Four-insights\u002Fpharmas-digital-rx-quantum-computing-in-drug-research-and-development",{"_key":60,"_type":9,"children":61,"markDefs":66,"style":17},"6a1d0730481e",[62],{"_key":63,"_type":13,"marks":64,"text":65},"2742fa7387a1",[],"Modelling different phenomena demands fundamentally different techniques. Characterizing a single material, for example, can involve chaining several computational methods together, and the number of possible combinations grows quickly as new techniques emerge. For scientists and machine learning researchers, this can create a laborious, clunky workflow: combining several complex software packages, spending hours debugging inconsistencies, and potentially using hundreds of lines of configuration and glue code.",[],{"_key":68,"_type":9,"children":69,"markDefs":87,"style":17},"66862ce43339",[70,74,79,83],{"_key":71,"_type":13,"marks":72,"text":73},"9c952a4b5812",[],"Today, we are excited to open-source ",{"_key":75,"_type":13,"marks":76,"text":78},"995b1089c661",[77,25],"f9efbde827a7","k",{"_key":80,"_type":13,"marks":81,"text":82},"05e36cb66b5a",[77],"UPS",{"_key":84,"_type":13,"marks":85,"text":86},"f246743b910e",[],", a molecular simulation engine for the AI era.",[88],{"_key":77,"_type":55,"internalSlug":56,"type":57,"url":89},"http:\u002F\u002Fgithub.com\u002Fcusp-ai-oss\u002Fkups",{"_key":91,"_type":9,"children":92,"markDefs":100,"style":17},"c74960b3ff0f",[93,96],{"_key":94,"_type":13,"marks":95,"text":78},"5f69d886ad4e",[25],{"_key":97,"_type":13,"marks":98,"text":99},"48bd76452300",[],"UPS is a toolkit that unifies diverse molecular simulation techniques behind a single composable interface. It gives computational scientists and ML researchers a shared, Python-native platform to run a broad range of simulations that are GPU-native and highly efficient, scaling seamlessly from a single laptop to data center GPUs.",[],{"_key":102,"_type":9,"children":103,"markDefs":115,"style":17},"2bb739e637a8",[104,108,111],{"_key":105,"_type":13,"marks":106,"text":107},"ef3ca43e5a82",[],"In practice, using ",{"_key":109,"_type":13,"marks":110,"text":78},"d024fa8c501f",[25],{"_key":112,"_type":13,"marks":113,"text":114},"f18fc33c4d79",[],"UPS, a researcher can write a custom simulation for a novel use case in an afternoon, rather than taking weeks of effort.",[],{"_key":117,"_type":118,"caption":119,"fit":120,"image":121,"width":132},"1ae710240c51","imageFigure","Proton hopping in sulfuric acid solution, simulated with kUPS. The larger green dots are oxygen, more faintly you can see bonds to hydrogen. Excess protons are shown in blue, leaving behind a fading trail of where they have been as they move throughout the H-bonding network via chemical reactions. The red pulses indicate the moment of proton-hopping.","natural",{"alt":122,"asset":123,"crop":56,"hotspot":56},"Animation of a molecular dynamics simulation showing proton hopping in a sulfuric acid solution. A network of green dots representing oxygen atoms…",{"_id":124,"metadata":125,"mimeType":130,"url":131},"image-35cce396b902cd355e3a6ae104177450d7b570ec-450x450-gif",{"dimensions":126,"lqip":129},{"aspectRatio":127,"height":128,"width":128},1,450,"data:image\u002Fjpeg;base64,\u002F9j\u002F2wBDAAYEBQYFBAYGBQYHBwYIChAKCgkJChQODwwQFxQYGBcUFhYaHSUfGhsjHBYWICwgIyYnKSopGR8tMC0oMCUoKSj\u002F2wBDAQcHBwoIChMKChMoGhYaKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCj\u002FwAARCAAUABQDASIAAhEBAxEB\u002F8QAGQABAQEAAwAAAAAAAAAAAAAAAAMCAQQI\u002F8QAIhAAAgIBAwQDAAAAAAAAAAAAAQIAAxESITEEEyIyBUFR\u002F8QAFwEAAwEAAAAAAAAAAAAAAAAAAAECA\u002F\u002FEABsRAQACAgMAAAAAAAAAAAAAAAABEQJREkFh\u002F9oADAMBAAIRAxEAPwDziaTWw7LhQR7E8zei3WosuCBhnc8CRaiwqrlSVYagR+SlgY1+LZRsaQfYyWVztLtkcpkHg45ibWi51BCvj6iOlX65BZKyquwU7YzOy1a0\u002FG09VXtcLMZiIYpz6RPV3gkiw+W5iIjuRxjT\u002F9k=","image\u002Fgif","https:\u002F\u002Fcdn.sanity.io\u002Fimages\u002Fho74vmlz\u002Fproduction\u002F35cce396b902cd355e3a6ae104177450d7b570ec-450x450.gif","column",{"_key":134,"_type":9,"children":135,"markDefs":140,"style":141,"anchorId":142},"be5198f17f68",[136],{"_key":137,"_type":13,"marks":138,"text":139},"fa8cef6b14f3",[],"Simulation tools weren’t built for the AI era",[],"h2","simulation-tools-werent-built-for-the-ai-era",{"_key":144,"_type":9,"children":145,"markDefs":150,"style":17},"251b3fcf2b53",[146],{"_key":147,"_type":13,"marks":148,"text":149},"673d1d9e0986",[],"Discovering materials requires combining hundreds of thousands of accurate molecular simulations with machine learning to select the most promising candidates for laboratory synthesis. High-throughput computational screening means exploring vast chemical spaces across thousands of simulation runs.",[],{"_key":152,"_type":9,"children":153,"markDefs":158,"style":17},"80eb55bf59af",[154],{"_key":155,"_type":13,"marks":156,"text":157},"aa3eb85eac59",[],"However, a one-size-fits-all approach fails when dealing with diverse phenomena. Researchers use different techniques including molecular dynamics for transport properties like diffusion, Monte Carlo sampling for adsorption equilibria and to predict gas uptake, and geometry optimization for predicting stable structures. Real-world applications rarely call for just one of these in isolation.",[],{"_key":160,"_type":9,"children":161,"markDefs":192,"style":17},"11332e31f24e",[162,166,171,175,180,183,188],{"_key":163,"_type":13,"marks":164,"text":165},"c110cbe28c32",[],"This means researchers juggle separate software packages for each method (",{"_key":167,"_type":13,"marks":168,"text":170},"86b985af811a",[169],"f725b25bceea","LAMMPS",{"_key":172,"_type":13,"marks":173,"text":174},"f5c79b2f12cf",[],", ",{"_key":176,"_type":13,"marks":177,"text":179},"c7e4a6b3e059",[178],"015ac32d0efd","GROMACS",{"_key":181,"_type":13,"marks":182,"text":174},"3b5cad90c6ce",[],{"_key":184,"_type":13,"marks":185,"text":187},"13c6ed69c30a",[186],"0f8590533ee2","RASPA",{"_key":189,"_type":13,"marks":190,"text":191},"d981b7a01114",[],", and others), each with its own input formats, data structures, and assumptions. This fragmentation is costly. Experimenting with a new method typically means writing complex C++ code from scratch for every framework.",[193,195,197],{"_key":169,"_type":55,"internalSlug":56,"type":57,"url":194},"https:\u002F\u002Fwww.lammps.org\u002F#gsc.tab=0",{"_key":178,"_type":55,"internalSlug":56,"type":57,"url":196},"https:\u002F\u002Fwww.gromacs.org\u002F",{"_key":186,"_type":55,"internalSlug":56,"type":57,"url":198},"https:\u002F\u002Firaspa.org\u002Fraspa\u002F",{"_key":200,"_type":9,"children":201,"markDefs":215,"style":17},"78f5e4746851",[202,206,211],{"_key":203,"_type":13,"marks":204,"text":205},"c555c1172a1c",[],"The rise of AI has made things harder still. Modern methods like machine-learning force fields, Boltzmann generators, and other learned components need to be deeply embedded in simulations for optimal performance, but traditional simulation packages were built to optimize well-known algorithms for CPUs, rather than GPUs. Their support for GPU-accelerated machine learning is bolted on through rigid interfaces that cannot keep pace with rapid developments in the field. Integrating a ",{"_key":207,"_type":13,"marks":208,"text":210},"9ae82f7d24b0",[209],"373fdf56ef51","PyTorch",{"_key":212,"_type":13,"marks":213,"text":214},"512ee5e70d71",[]," model into a C++ engine means implementing custom bridges, cross-language memory management, and sacrificing the flexibility that made the model easy to develop in the first place.",[216],{"_key":209,"_type":55,"internalSlug":56,"type":57,"url":217},"https:\u002F\u002Fgithub.com\u002Fpytorch\u002Fpytorch",{"_key":219,"_type":9,"children":220,"markDefs":232,"style":17},"76b59abb4ffa",[221,225,228],{"_key":222,"_type":13,"marks":223,"text":224},"03cea9c50307",[],"Meanwhile, the tools powering AI development (Python, automatic differentiation, GPU-native computation) remain foreign to traditional simulation codes. We built ",{"_key":226,"_type":13,"marks":227,"text":78},"1cab052763b6",[25],{"_key":229,"_type":13,"marks":230,"text":231},"aec6f7810618",[],"UPS to close that gap.",[],{"_key":234,"_type":9,"children":235,"markDefs":240,"style":141,"anchorId":241},"7e12c42d422b",[236],{"_key":237,"_type":13,"marks":238,"text":239},"5662529da4ba",[],"A modern foundation for molecular simulations",[],"a-modern-foundation-for-molecular-simulations",{"_key":243,"_type":9,"children":244,"markDefs":252,"style":17},"4f554ab68ab8",[245,248],{"_key":246,"_type":13,"marks":247,"text":78},"cb3a3d96db55",[25],{"_key":249,"_type":13,"marks":250,"text":251},"e321bf0251b5",[],"UPS is built on three design principles: composability, structured data, and batched execution. Every operation flows through JAX, so simulations run natively on CPU, GPU, and TPU with no code changes.",[],{"_key":254,"_type":9,"children":255,"markDefs":269,"style":17},"2aff28e6f20c",[256,260,265],{"_key":257,"_type":13,"marks":258,"text":259},"cce00097ffb4",[],"Working in collaboration with NVIDIA, we have optimized for GPU performance in particular, achieving up to 49× throughput over widely adopted tools like RASPA for certain simulations. And, because many models today live in PyTorch, we also open-source ",{"_key":261,"_type":13,"marks":262,"text":264},"945c4d6eb6ca",[263],"b8833fa75b0d","Tojax",{"_key":266,"_type":13,"marks":267,"text":268},"e80aa90cbfc2",[],", a library for transforming PyTorch functions into JAX, meeting developers where they are.",[270],{"_key":263,"_type":55,"internalSlug":56,"type":57,"url":271},"http:\u002F\u002Fgithub.com\u002Fcusp-ai-oss\u002Ftojax",{"_key":273,"_type":9,"children":274,"markDefs":288,"style":17},"b0d014bf90ec",[275,279,284],{"_key":276,"_type":13,"marks":277,"text":278},"6e2a970cc897",[],"This work is in addition to our ongoing collaboration with NVIDIA on ",{"_key":280,"_type":13,"marks":281,"text":283},"777d64d7f604",[282],"e9208822e28e","ALCHEMI",{"_key":285,"_type":13,"marks":286,"text":287},"a089728db2df",[],", their platform for accelerating AI-driven simulation in chemistry and materials science.",[289],{"_key":282,"_type":55,"internalSlug":56,"type":57,"url":290},"https:\u002F\u002Fdeveloper.nvidia.com\u002Fcuda\u002Fcuda-x-libraries\u002Falchemi",{"_key":292,"_type":293,"attribution":294,"quote":295,"role":296},"8b64c5962a91","pullQuote","Ian Buck","We’ve been working closely with the CuspAI team to help push the boundaries of GPU-accelerated molecular simulation. The ALCHEMI collaboration with CuspAI demonstrates what’s possible when modern simulation software is built natively for accelerated computing. This approach accelerates scientific discovery in unprecedented ways, unlocking the kind of throughput gains that let researchers move from thousands of simulations to millions, and thus from hypothesis to discovery, faster than ever before.","Vice President of Hyperscale and HPC Computing, NVIDIA",{"_key":298,"_type":9,"children":299,"markDefs":304,"style":141,"anchorId":305},"43e70eeaa3d0",[300],{"_key":301,"_type":13,"marks":302,"text":303},"eb276d5a5945",[],"Composability",[],"composability",{"_key":307,"_type":9,"children":308,"markDefs":320,"style":17},"6a72a8416d5a",[309,313,316],{"_key":310,"_type":13,"marks":311,"text":312},"faccfea6a78a",[],"To fully characterize a material, scientists often need to link multiple techniques into a single workflow. In traditional tools, each combination means writing new glue code from scratch. ",{"_key":314,"_type":13,"marks":315,"text":78},"308d9983a0bd",[25],{"_key":317,"_type":13,"marks":318,"text":319},"b0d74ad32dd5",[],"UPS is designed so that any method can be combined with any other by construction.",[],{"_key":322,"_type":9,"children":323,"markDefs":328,"style":17},"b141a26d7f15",[324],{"_key":325,"_type":13,"marks":326,"text":327},"5766cf565ce3",[],"Every operation follows a single abstraction: propagate a state to its next state. The next state depends only on the current state and a random key. This has practical consequences. Simulations are bitwise reproducible, because replaying the same sequence of keys from a given state always yields identical trajectories. Checkpointing is trivial, since the state is a complete snapshot with no hidden history, and any propagator can be tested in isolation by feeding it a single state.",[],{"_key":330,"_type":9,"children":331,"markDefs":336,"style":17},"828a63b15878",[332],{"_key":333,"_type":13,"marks":334,"text":335},"ee7cce3b14b9",[],"For scientists, this means fewer hours spent debugging subtle nondeterminism, and more confidence that results will hold up when shared, reproduced, or scaled.",[],{"_key":338,"_type":118,"fit":120,"image":339,"width":132},"bdaf2c3b340b",{"alt":340,"asset":341,"crop":56,"hotspot":56},"A flow diagram of the kUPS propagator interface. A grey rounded rectangle labelled ‘Input’ contains two smaller boxes: ‘Random Key’ and ‘State’. Two…",{"_id":342,"metadata":343,"mimeType":349,"url":350},"image-72110b70bac04dee0959773ba513f61f7905fab0-1394x579-png",{"dimensions":344,"lqip":348},{"aspectRatio":345,"height":346,"width":347},2.407599309153713,579,1394,"data:image\u002Fpng;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAICAYAAAD5nd\u002FtAAAACXBIWXMAAAPoAAAD6AG1e1JrAAABNUlEQVQokYVSy04CQRDkQ41fYPwao1f4AD3rlRPh5lkWWJLdGfYpzIOdR0+Z2V00xAQ6qUzPoasfVRPnHIQQaJoGdV33aNsWSikQEWJQAGIaEXA9Jlpr5HmOzWaD1WqFJEn6vCgKGGNAFCBOAew7oDoGGHeDUAiBLMuw2+0uwBhDbHbQHosvh+nc4HVpsd4TrAvwFGA9+gbWAX6cfiKl7IsjON+Dc44sz\u002Fu\u002FkBJr3uHpQ+PuWeFhqvH+adBKB2UcWkkoj4RaEpSJTcYJY3FcsaoqlGUJxniPSJiwgfD+ReFxNhA20kJ1VwjzcaIzUTwB4xz6dMJBDSvP5gZvS4tkTzCO4IhgfUDnwuXKZ1G22+2IFGmaoigvReGjKJ0djh\u002FCoPf4\u002FInivf+1zRn\u002FbDNaprfNDd\u002F8ACPnYZfzZ+51AAAAAElFTkSuQmCC","image\u002Fpng","https:\u002F\u002Fcdn.sanity.io\u002Fimages\u002Fho74vmlz\u002Fproduction\u002F72110b70bac04dee0959773ba513f61f7905fab0-1394x579.png",{"_key":352,"_type":9,"children":353,"markDefs":365,"style":17},"335342b5e3e7",[354,358,361],{"_key":355,"_type":13,"marks":356,"text":357},"f6f691b9e1cd",[],"The shared interface is what makes it all composable. In ",{"_key":359,"_type":13,"marks":360,"text":78},"b73fd4a76d13",[25],{"_key":362,"_type":13,"marks":363,"text":364},"51756a08d1f6",[],"UPS, an integrator, a thermostat, a Monte Carlo move, a force field evaluation are all propagators. Because they share the same signature, they snap together freely. A new potential works with every existing sampler, and a new sampling method works with every existing potential.",[],{"_key":367,"_type":9,"children":368,"markDefs":373,"style":17},"f6062351299c",[369],{"_key":370,"_type":13,"marks":371,"text":372},"a61b7f47ef14",[],"This extends naturally to learned components: a machine-learning force field, a generative model serving as the proposal distribution in a Markov chain Monte Carlo (MCMC) simulation, or a learned large-timestep integrator all plug in through the same abstraction. This could reduce the timeframe for wiring a custom simulation for a novel use case from weeks, to a single afternoon.",[],{"_key":375,"_type":118,"caption":376,"fit":120,"image":377,"width":132},"c2fc8b1be835","All three simulations shown above hold particle count (N) constant. NVE (constant volume and energy) chains a potential and an integrator; adding a thermostat gives NVT (constant volume and temperature); adding a barostat on top gives NPT (constant pressure and temperature). Each ensemble is assembled from the same reusable propagators — simply composed in a different order.",{"alt":378,"asset":379,"crop":56,"hotspot":56},"A flow diagram with three rows showing how kUPS composes reusable propagators into different ensembles. NVE chains Potential and Integrator. NVT…",{"_id":380,"metadata":381,"mimeType":349,"url":387},"image-8f93eae2817584dac8becf02dffafc694f873bb8-1611x544-png",{"dimensions":382,"lqip":386},{"aspectRatio":383,"height":384,"width":385},2.9613970588235294,544,1611,"data:image\u002Fpng;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAYAAAAIy204AAAACXBIWXMAAAPoAAAD6AG1e1JrAAABCklEQVQokWWR2W7EIAxF5\u002F9\u002FNZ1AQli9nAqSalrV0n0A44N9\u002FXJ3RGTJzPiEYzhd4epGGUqVSh0VMVn5W3\u002FjNSGttSVVvW\u002FdcJN1Ts3YLiWUTqwnoQTqKJgNMAX3\u002F8De+wKaPg+04+1AayLVvoB7GYR6EMpOaSfWIt5O0HHXPOBPh7Wio4E0vF1Y2tD0JuXGdglhAsvBnndKiWjasPSF97wa4LHh02G50Lxj+Y2miOyBEQ7O6+7wNzBfx8pLiFiOWHmviSb0ZfYspVWszOSEJiReyFnIVYh1eimkljnbSa0ZOW5ZSatmjW\u002FKa+3AHZ8bVlmeuAouU4qqM9SR+bHp2vD0euaWnpqfkb8BElIkcLeTm9gAAAAASUVORK5CYII=","https:\u002F\u002Fcdn.sanity.io\u002Fimages\u002Fho74vmlz\u002Fproduction\u002F8f93eae2817584dac8becf02dffafc694f873bb8-1611x544.png",{"_key":389,"_type":9,"children":390,"markDefs":395,"style":141,"anchorId":396},"c1db92eeca79",[391],{"_key":392,"_type":13,"marks":393,"text":394},"ec1779d9e624",[],"Data structure",[],"data-structure",{"_key":398,"_type":9,"children":399,"markDefs":404,"style":17},"b9c030199728",[400],{"_key":401,"_type":13,"marks":402,"text":403},"03684b19b2e4",[],"Molecular simulations involve data at different scales: per-atom positions and forces, per-molecule charges and topologies, per-system temperatures and pressures. Organizing such data in memory and ensuring their efficient use is a challenge for GPUs. When thousands of threads need to read positions simultaneously, scattering that data across separate objects means scattered memory access and idle hardware.",[],{"_key":406,"_type":118,"fit":120,"image":407,"width":132},"2226e75fcbf3",{"alt":408,"asset":409,"crop":56,"hotspot":56},"Diagram of kUPS’s columnar Table data layout. Three 3D isometric grids are shown. Top right: a Systems table with M rows and columns for Key,…",{"_id":410,"metadata":411,"mimeType":349,"url":417},"image-d36c7c4202c27abe651b8ef52354a98ae3979f64-2340x1378-png",{"dimensions":412,"lqip":416},{"aspectRatio":413,"height":414,"width":415},1.6981132075471699,1378,2340,"data:image\u002Fpng;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAMCAYAAABiDJ37AAAACXBIWXMAAAPoAAAD6AG1e1JrAAACX0lEQVQokY2Ty09TQRSH+2+6MXFjXMnOaOLKlQsxMSAUCAaMRlEaQDFRKIhCxFpeBQqUvmlL20tft7f3zsx9feZCJOhGTnImZzHzze935kwIwPd9hFQ0OxblM0m6rJOpN9F6LYR1his6+K7NdSIULMp2KJZrLP7MMLncZnChxUgszUImRi33GVmN40o9uPn\u002FQN\u002F3sYRkez\u002FNwNsYd8dq3BxrcPvDJkPxCNmNZ1jHMzimhu97gZ\u002FrATd3j3n6ap07w1VuhJvcmo0zEHtH+tcTzFQEp1e\u002FPlBKRSpb4n00Sf9cg0dzHR6vHRI5WKJ88AZx8g1HtOEKMDj3b1720HVdOrpBvtJir2CyXRAkqk1yzRO6zTS2cYrnKv6E53nYysYyJT1DYJoS23bOoefAoAiggVKjJ+iZEqlslKNwHIXnOZfKXM\u002FD6FnUTjtk93WSsQ6ZHQ29EkyDIOR5PlIpej1JsybIH3U5SXUwtC6ekOdqrtoyLcFRusTKYpZPL1pM9zdYHkpwuvADVS4TCuavVKlzkKywEe3yZcJgaaJEMbqLyOWxLRNl2yilsG2bVkfn+\u002FoekyObjD88Y7ivzsy9JQqjU1gHSUKB\u002FGQqx9rqMctTbWYHTKLhHIXpFYydBJauY1lBnyyEkDRabb6u7jA+GGf0gUa4r87H+1GK4deY+\u002FuEArtVrUE+r5HZNUnFJaVEi+5hFlWt4kiJ4zgX6brohsFWIs185JC5cJv55202Xm5xNr+IzOcvxiZ4IaVspHBRwsORLq5S+I7z1+8IqmCfprXJZRrnAgpJQTNdxyoUcbtdfgMz3Xg4vvV8XwAAAABJRU5ErkJggg==","https:\u002F\u002Fcdn.sanity.io\u002Fimages\u002Fho74vmlz\u002Fproduction\u002Fd36c7c4202c27abe651b8ef52354a98ae3979f64-2340x1378.png",{"_key":419,"_type":9,"children":420,"markDefs":428,"style":17},"270ccc97fecd",[421,424],{"_key":422,"_type":13,"marks":423,"text":78},"80bbf3207d6b",[25],{"_key":425,"_type":13,"marks":426,"text":427},"242b3110b2d5",[],"UPS stores simulation data in Tables, columnar containers where every property shares the same leading dimension. Each row carries a unique key, analogous to a primary key in a SQL database, and tables reference each other through index arrays that act as foreign keys. This gives us structure-of-arrays layout (one contiguous array per property) rather than arrays-of-structures (one object per atom). With many checks and strict type checking, this significantly reduces implementation errors and the manual work of remembering data layouts when implementing new simulations. On GPUs, where performance depends on coalesced memory access across thousands of threads, this layout is the difference between saturating the hardware and wasting it.",[],{"_key":430,"_type":9,"children":431,"markDefs":436,"style":141,"anchorId":437},"8c06c1955480",[432],{"_key":433,"_type":13,"marks":434,"text":435},"6e0695701671",[],"Performance",[],"performance",{"_key":439,"_type":9,"children":440,"markDefs":452,"style":17},"e56b44dfc6db",[441,445,448],{"_key":442,"_type":13,"marks":443,"text":444},"41c229934f1e",[],"Traditional simulation tools process one system at a time, leaving most of a GPU’s capacity idle. ",{"_key":446,"_type":13,"marks":447,"text":78},"4738e62460c1",[25],{"_key":449,"_type":13,"marks":450,"text":451},"65c645f916ed",[],"UPS is built around batching: running many independent systems in parallel as a single vectorized computation. The columnar table layout makes this natural: stacking systems along the leading dimension turns a batch of simulations into a single dense array operation. This is the backbone of our high-throughput workflows and where the performance gains come from.",[],{"_key":454,"_type":118,"fit":120,"image":455,"width":132},"bf0b1e2c642c",{"alt":456,"asset":457,"crop":56,"hotspot":56},"Two benchmark graphs, the first showing 1.8x throughput for kUPS vs. ASE for an NVT molecular dynamics simulation. The second shows 49x throughput…",{"_id":458,"metadata":459,"mimeType":349,"url":465},"image-32f5e28c656ade2c1520a43de64b5f2ac7191dd1-2183x680-png",{"dimensions":460,"lqip":464},{"aspectRatio":461,"height":462,"width":463},3.210294117647059,680,2183,"data:image\u002Fpng;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAGCAYAAADDl76dAAAACXBIWXMAAAPoAAAD6AG1e1JrAAABLElEQVQYlVWQ207dMBBF8\u002F8fVZW+VBW0QJEoV1E4FQmNjy\u002FxOHac2KtKDqIw0n4YzdaaPdMsy4IEwRiNNhrnLDFGSim8r7UfY8RYg7EWESHPGagffM08zwegtVjnEAnknKn1I3BdnFJEgieMQkxxA9bN9x\u002Fa+OB5MS1P\u002FQM79UDndujY4fOeXBKV+ppuxMoeFZ5R8Zl9bNGpxU09aQmU1wCNMn+57X9y1h3xo\u002FvEef+ZC\u002F2Fe3eKm9RmXNMN4tiZW37pr9v8oCOuzDf68Q9zmbbljURPJ4\u002F8Hi64s6fc6BOu9TGP7hI\u002FGWqtm1KOqNDy5C+5H864Md+50sfcmXP02LGU+QBcE+QykRbBR4OWHjW0GFFMc3r7zQqdS97Ok8lhgkINHXv\u002FQpzk7ef\u002FAFV7y3QnHx7lAAAAAElFTkSuQmCC","https:\u002F\u002Fcdn.sanity.io\u002Fimages\u002Fho74vmlz\u002Fproduction\u002F32f5e28c656ade2c1520a43de64b5f2ac7191dd1-2183x680.png",{"_key":467,"_type":9,"children":468,"markDefs":473,"style":17},"768d3d911ea5",[469],{"_key":470,"_type":13,"marks":471,"text":472},"b2452cc88bc1",[],"As an example, we benchmark two representative workloads on a single NVIDIA L4 GPU, using metal-organic frameworks — porous materials with downstream uses including carbon capture — as a test case. We are actively expanding our suite of benchmarks to include further examples across diverse material classes.",[],{"_key":475,"_type":9,"children":476,"markDefs":506,"style":17},"483d589b7d66",[477,481,486,490,493,497,502],{"_key":478,"_type":13,"marks":479,"text":480},"08f361156845",[],"First, we explore NVT molecular dynamics of a single 114-atom metal–organic framework (MOF) using the ",{"_key":482,"_type":13,"marks":483,"text":485},"bdb821ec8024",[484],"d67dd98fcd2a","MACE",{"_key":487,"_type":13,"marks":488,"text":489},"dac40b1a33d0",[]," neural network potential. Even without batching, ",{"_key":491,"_type":13,"marks":492,"text":78},"c7141eef91d7",[25],{"_key":494,"_type":13,"marks":495,"text":496},"cba4648fe8c3",[],"UPS achieves 1.8× the throughput of ",{"_key":498,"_type":13,"marks":499,"text":501},"c7827af613b6",[500],"483871c8f55e","ASE",{"_key":503,"_type":13,"marks":504,"text":505},"71237a1bf64f",[],", the standard Python interface for running MACE, purely from lower overhead in the simulation loop. Batching multiple systems would increase performance further by saturating the GPU across independent runs.",[507,509],{"_key":484,"_type":55,"internalSlug":56,"type":57,"url":508},"https:\u002F\u002Fgithub.com\u002FACEsuit\u002Fmace",{"_key":500,"_type":55,"internalSlug":56,"type":57,"url":510},"https:\u002F\u002Fase-lib.org\u002F",{"_key":512,"_type":9,"children":513,"markDefs":532,"style":17},"3fc0e0802975",[514,518,521,525,528],{"_key":515,"_type":13,"marks":516,"text":517},"495f6e67e208",[],"Second, we look at grand-canonical Monte Carlo (µVT) for CO₂ adsorption in a 3×3×3 supercell of the same framework (3,078 framework atoms, ~30 CO₂ molecules at average loading). RASPA, the most widely used tool for this task, runs on a single CPU core. ",{"_key":519,"_type":13,"marks":520,"text":78},"bd423cfdb5ea",[25],{"_key":522,"_type":13,"marks":523,"text":524},"4e1f77270da2",[],"UPS replaces it with a fully vectorized GPU implementation, achieving 49× throughput on the same workload. Using ",{"_key":526,"_type":13,"marks":527,"text":78},"6a96bdc6a40f",[25],{"_key":529,"_type":13,"marks":530,"text":531},"25aea9532914",[],"UPS, an entire class of simulations that was architecturally locked to serial CPU execution can now run natively on accelerators.",[],{"_key":534,"_type":9,"children":535,"markDefs":540,"style":141,"anchorId":541},"b9ed51887265",[536],{"_key":537,"_type":13,"marks":538,"text":539},"f46674a9b7f0",[],"Putting it together",[],"putting-it-together",{"_key":543,"_type":9,"children":544,"markDefs":556,"style":17},"8e4a2f4bf8d3",[545,549,552],{"_key":546,"_type":13,"marks":547,"text":548},"3647e70ffae9",[],"In a traditional workflow, setting up a molecular dynamics simulation with a machine-learning potential might involve hundreds of lines of configuration spread across multiple files and languages, plus careful glue code to connect the model to the simulation engine. With ",{"_key":550,"_type":13,"marks":551,"text":78},"437ece82a6f6",[25],{"_key":553,"_type":13,"marks":554,"text":555},"926e30a638cf",[],"UPS, the same simulation takes a few lines of Python. Users can define a data layout, wire up a potential and integrator, and run:",[],{"_key":558,"_type":559,"code":560,"html":564},"4fc0be68b37c","codeBlock",{"_type":561,"code":562,"language":563},"code","class MyAtoms:\n    positions: Array\n    atomic_numbers: Array\n    ...\n\nclass MySystems:\n    unitcell: UnitCell\n    ...\n\nclass MyFirstState:\n    atoms: Table[ParticleId, MyAtoms]\n    systems: Table[SystemId, MySystems]\n    mace: MACEModel\n    ...\n\natoms = MyAtoms(...) # populate atoms\nsystems = MySystem(...) # populate systems\nstate = MyFirstState(atoms, systems, ...) # populate state\nchain = key_chain(jax.random.key(0)) # random key chain\n\npotential = make_tojaxed_potential_from_state(\n    identity_lens(MyFirstState), # lens: tells the potential where to find its inputs in the state\n    compute_position_and_unitcell_gradients=True\n)\n\npropagator = make_md_propagator(state_lens, \"verlet\", potential)\n\nfor _ in range(1000):\n    state = propagator(next(chain), state)","python","\u003Cpre class=\"shiki cusp\" style=\"background-color:#262626;color:#E5E5E5\" tabindex=\"0\">\u003Ccode>\u003Cspan class=\"line\">\u003Cspan style=\"color:#88FF82\">class\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\"> MyAtoms\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">:\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#E5E5E5\">    positions\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">:\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\"> Array\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#E5E5E5\">    atomic_numbers\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">:\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\"> Array\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#7DD4E0\">    ...\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#88FF82\">class\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\"> MySystems\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">:\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#E5E5E5\">    unitcell\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">:\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\"> UnitCell\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#7DD4E0\">    ...\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#88FF82\">class\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\"> MyFirstState\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">:\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#E5E5E5\">    atoms\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">:\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\"> Table\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">[\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\">ParticleId\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">,\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\"> MyAtoms\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">]\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#E5E5E5\">    systems\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">:\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\"> Table\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">[\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\">SystemId\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">,\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\"> MySystems\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">]\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#E5E5E5\">    mace\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">:\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\"> MACEModel\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#7DD4E0\">    ...\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#E5E5E5\">atoms \u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">=\u003C\u002Fspan>\u003Cspan style=\"color:#FAFAFA\"> MyAtoms\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">(\u003C\u002Fspan>\u003Cspan style=\"color:#7DD4E0\">...\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">)\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\"> # populate atoms\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#E5E5E5\">systems \u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">=\u003C\u002Fspan>\u003Cspan style=\"color:#FAFAFA\"> MySystem\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">(\u003C\u002Fspan>\u003Cspan style=\"color:#7DD4E0\">...\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">)\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\"> # populate systems\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#E5E5E5\">state \u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">=\u003C\u002Fspan>\u003Cspan style=\"color:#FAFAFA\"> MyFirstState\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">(\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\">atoms\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">,\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\"> systems\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">,\u003C\u002Fspan>\u003Cspan style=\"color:#7DD4E0\"> ...\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">)\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\"> # populate state\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#E5E5E5\">chain \u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">=\u003C\u002Fspan>\u003Cspan style=\"color:#FAFAFA\"> key_chain\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">(\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\">jax\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">.\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\">random\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">.\u003C\u002Fspan>\u003Cspan style=\"color:#FAFAFA\">key\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">(\u003C\u002Fspan>\u003Cspan style=\"color:#7DD4E0\">0\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">))\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\"> # random key chain\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#E5E5E5\">potential \u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">=\u003C\u002Fspan>\u003Cspan style=\"color:#FAFAFA\"> make_tojaxed_potential_from_state\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">(\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#FAFAFA\">    identity_lens\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">(\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\">MyFirstState\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">),\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\"> # lens: tells the potential where to find its inputs in the state\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#E5E5E5\">    compute_position_and_unitcell_gradients\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">=\u003C\u002Fspan>\u003Cspan style=\"color:#88FF82\">True\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#A1A1A1\">)\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#E5E5E5\">propagator \u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">=\u003C\u002Fspan>\u003Cspan style=\"color:#FAFAFA\"> make_md_propagator\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">(\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\">state_lens\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">,\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\"> \"\u003C\u002Fspan>\u003Cspan style=\"color:#7DD4E0\">verlet\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">\"\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">,\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\"> potential\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">)\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#88FF82\">for\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\"> _ \u003C\u002Fspan>\u003Cspan style=\"color:#88FF82\">in\u003C\u002Fspan>\u003Cspan style=\"color:#88FF82\"> range\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">(\u003C\u002Fspan>\u003Cspan style=\"color:#7DD4E0\">1000\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">):\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#E5E5E5\">    state \u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">=\u003C\u002Fspan>\u003Cspan style=\"color:#FAFAFA\"> propagator\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">(\u003C\u002Fspan>\u003Cspan style=\"color:#88FF82\">next\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">(\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\">chain\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">),\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\"> state\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">)\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fcode>\u003C\u002Fpre>",{"_key":566,"_type":9,"children":567,"markDefs":581,"style":17},"9ba5c16ee878",[568,572,577],{"_key":569,"_type":13,"marks":570,"text":571},"6c7be4657c7e",[],"Swapping MACE for ",{"_key":573,"_type":13,"marks":574,"text":576},"957807779e59",[575],"3e04387975ed","UMA",{"_key":578,"_type":13,"marks":579,"text":580},"1fd6d770b74b",[]," is a single line change, adding a Monte Carlo move alongside the dynamics — or inserting a learned large-timestep model between integration steps — takes just a few lines more. With traditional tools, changes like these would typically mean switching to an entirely different package and rewriting the simulation from scratch.",[582],{"_key":575,"_type":55,"internalSlug":56,"type":57,"url":583},"https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.23971",{"_key":585,"_type":9,"children":586,"markDefs":591,"style":141,"anchorId":592},"ecdb832cdc22",[587],{"_key":588,"_type":13,"marks":589,"text":590},"d0479bd17118",[],"Methods and potentials",[],"methods-and-potentials",{"_key":594,"_type":9,"children":595,"markDefs":603,"style":17},"2f9f72c5ad7c",[596,599],{"_key":597,"_type":13,"marks":598,"text":78},"0ea737334209",[25],{"_key":600,"_type":13,"marks":601,"text":602},"a8d07c7d08e8",[],"UPS ships with a broad set of methods and potentials covering the most common molecular simulation tasks, all built using the same composable abstractions available to any user.",[],{"_key":605,"_type":9,"children":606,"markDefs":618,"style":17},"bd59f61e4596",[607,611,614],{"_key":608,"_type":13,"marks":609,"text":610},"7e9a44e55735",[],"On the sampling side, this includes molecular dynamics ensembles (NVE, NVT, NPT), Monte Carlo methods (NVT, µVT), and geometry optimizers (FIRE, L-BFGS). For interactions, ",{"_key":612,"_type":13,"marks":613,"text":78},"8ed4f640a00c",[25],{"_key":615,"_type":13,"marks":616,"text":617},"c27db24d2ea5",[],"UPS provides classical potentials such as Lennard-Jones, Coulomb (via Ewald summation), harmonic bonds and angles, and Morse, alongside ML force fields including MACE and UMA. Every built-in method was written with the same public API you would use to build your own, and anything you add will work with every existing component from day one.",[],{"_key":620,"_type":9,"children":621,"markDefs":626,"style":141,"anchorId":627},"7a37333b0f60",[622],{"_key":623,"_type":13,"marks":624,"text":625},"4327c4dedc22",[],"Bridging the gap: Tojax",[],"bridging-the-gap-tojax",{"_key":629,"_type":9,"children":630,"markDefs":643,"style":17},"78a3d2817b48",[631,635,639],{"_key":632,"_type":13,"marks":633,"text":634},"eb9669290205",[],"Rather than asking users to rewrite their PyTorch models, we additionally open-source ",{"_key":636,"_type":13,"marks":637,"text":264},"38d8a47f33a7",[638],"b7f1b2c007b9",{"_key":640,"_type":13,"marks":641,"text":642},"f8cbd905cfb4",[],", a library that converts any PyTorch function into a JAX function with a single line of code.",[644],{"_key":638,"_type":55,"internalSlug":56,"type":57,"url":645},"https:\u002F\u002Fgithub.com\u002Fcusp-ai-oss\u002Ftojax",{"_key":647,"_type":559,"code":648,"html":650},"475c350a5233",{"_type":561,"code":649,"language":563},"model = SimpleModel() # any PyTorch nn.Module\njax_model = tojax(model) # → callable in JAX\n\nx = jnp.ones((128, 32, 10))\noutput = jax.jit(jax.vmap(jax_model))(x) # JAX transformation compatible","\u003Cpre class=\"shiki cusp\" style=\"background-color:#262626;color:#E5E5E5\" tabindex=\"0\">\u003Ccode>\u003Cspan class=\"line\">\u003Cspan style=\"color:#E5E5E5\">model \u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">=\u003C\u002Fspan>\u003Cspan style=\"color:#FAFAFA\"> SimpleModel\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">()\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\"> # any PyTorch nn.Module\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#E5E5E5\">jax_model \u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">=\u003C\u002Fspan>\u003Cspan style=\"color:#FAFAFA\"> tojax\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">(\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\">model\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">)\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\"> # → callable in JAX\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#E5E5E5\">x \u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">=\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\"> jnp\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">.\u003C\u002Fspan>\u003Cspan style=\"color:#FAFAFA\">ones\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">((\u003C\u002Fspan>\u003Cspan style=\"color:#7DD4E0\">128\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">,\u003C\u002Fspan>\u003Cspan style=\"color:#7DD4E0\"> 32\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">,\u003C\u002Fspan>\u003Cspan style=\"color:#7DD4E0\"> 10\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">))\u003C\u002Fspan>\u003C\u002Fspan>\n\u003Cspan class=\"line\">\u003Cspan style=\"color:#E5E5E5\">output \u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">=\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\"> jax\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">.\u003C\u002Fspan>\u003Cspan style=\"color:#FAFAFA\">jit\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">(\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\">jax\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">.\u003C\u002Fspan>\u003Cspan style=\"color:#FAFAFA\">vmap\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">(\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\">jax_model\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">))(\u003C\u002Fspan>\u003Cspan style=\"color:#E5E5E5\">x\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\">)\u003C\u002Fspan>\u003Cspan style=\"color:#A1A1A1\"> # JAX transformation compatible\u003C\u002Fspan>\u003C\u002Fspan>\u003C\u002Fcode>\u003C\u002Fpre>",{"_key":652,"_type":9,"children":653,"markDefs":681,"style":17},"4c30f787a33b",[654,658,663,667,670,674,677],{"_key":655,"_type":13,"marks":656,"text":657},"b19cf7f0ab25",[],"A pre-trained MACE, ",{"_key":659,"_type":13,"marks":660,"text":662},"3b3fb2374cd2",[661],"8e9173fabd03","NequIP",{"_key":664,"_type":13,"marks":665,"text":666},"6ee024b686ef",[],", or custom PyTorch potential can be brought into ",{"_key":668,"_type":13,"marks":669,"text":78},"da20c19c6d31",[25],{"_key":671,"_type":13,"marks":672,"text":673},"78273ded6783",[],"UPS with minimal code changes. With Tojax, the model composes with jit, vmap, grad, and all other JAX transformations, the same ones ",{"_key":675,"_type":13,"marks":676,"text":78},"1b04ab7410f5",[25],{"_key":678,"_type":13,"marks":679,"text":680},"4ffceaba2cdc",[],"UPS relies on for batching and automatic differentiation.",[682],{"_key":661,"_type":55,"internalSlug":56,"type":57,"url":683},"https:\u002F\u002Fgithub.com\u002Fmir-group\u002Fnequip",{"_key":685,"_type":9,"children":686,"markDefs":691,"style":17},"9aa3e708d15f",[687],{"_key":688,"_type":13,"marks":689,"text":690},"3822237290a2",[],"There is no performance penalty: Tojax traces through the PyTorch module and emits a native JAX computation graph, compiled and optimized end-to-end by XLA with no Python-level back-and-forth between frameworks at runtime.",[],{"_key":693,"_type":9,"children":694,"markDefs":699,"style":141,"anchorId":700},"4362739bebf1",[695],{"_key":696,"_type":13,"marks":697,"text":698},"af69bf63db02",[],"Get started",[],"get-started",{"_key":702,"_type":9,"children":703,"markDefs":723,"style":17},"c0eda6ad0fb0",[704,708,711,715,719],{"_key":705,"_type":13,"marks":706,"text":78},"db1b9c967bca",[707,25],"463a1fb7f911",{"_key":709,"_type":13,"marks":710,"text":82},"fc9d73825c3d",[707],{"_key":712,"_type":13,"marks":713,"text":714},"770df2d2c869",[]," and ",{"_key":716,"_type":13,"marks":717,"text":264},"ecf486f3bc08",[718],"9faa5c4605b5",{"_key":720,"_type":13,"marks":721,"text":722},"b1daa1c6f426",[]," are open-source. We built them because we needed a simulation toolkit that could keep up with the pace of AI research: one where trying a new idea means writing a few lines of Python, not weeks of C++. One where simulations run natively on GPUs without compromise, and where classical methods, machine-learning potentials, and generative models all live within a single composable framework.",[724,725],{"_key":707,"_type":55,"internalSlug":56,"type":57,"url":89},{"_key":718,"_type":55,"internalSlug":56,"type":57,"url":271},{"_key":727,"_type":9,"children":728,"markDefs":733,"style":36},"22d855dab5ce",[729],{"_key":730,"_type":13,"marks":731,"text":732},"80885ea87189",[],"Whether scientists and machine learning researchers are screening millions of materials for carbon capture, semiconductors or catalysts, exploring drug-protein interactions, or developing entirely new simulation methods, we hope kUPS will give them the foundation to move faster.",[],{"_key":735,"_type":9,"children":736,"markDefs":751,"style":17},"0c771212e900",[737,742,746],{"_key":738,"_type":13,"marks":739,"text":741},"e618e427620c",[740],"5dbede5e48f2","github.com\u002Fcusp-ai-oss\u002Fkups",{"_key":743,"_type":13,"marks":744,"text":745},"df61e83f4c75",[],"\n",{"_key":747,"_type":13,"marks":748,"text":750},"adc083b5a0ea",[749],"dcbd40258140","github.com\u002Fcusp-ai-oss\u002Ftojax",[752,753],{"_key":740,"_type":55,"internalSlug":56,"type":57,"url":89},{"_key":749,"_type":55,"internalSlug":56,"type":57,"url":271},"Nicholas Gao, Jonas Köhler, Felix Hanke, Anita Ramanan",{"caption":756,"credit":56,"gif":56,"image":757,"panelColor":767,"type":768,"video":56},"CuspAI open-sources kUPS, a molecular simulation engine for the AI era",{"alt":756,"asset":758,"crop":56,"hotspot":56},{"_id":759,"metadata":760,"mimeType":349,"url":766},"image-6b92362855c51347bc1e9531279166ee819079d7-2168x1080-png",{"dimensions":761,"lqip":765},{"aspectRatio":762,"height":763,"width":764},2.0074074074074075,1080,2168,"data:image\u002Fpng;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAKCAYAAAC0VX7mAAAACXBIWXMAAAPoAAAD6AG1e1JrAAABhklEQVQokY1S20pCURA9nxcFFT1ID10oK5IuRAaVVOrMbM1f6kGCIoiI6CISWUlEEWGBdvG+YvY+xkkiehjO7Dlrz8xae3mpJoFqBNMkbMOPdiD\u002FK9qEVIOwWSGsvzCkRvCkQYi\u002FM7juQGkflGr9r2m6RaAqIfHO9p4ndZ3AdksF6JTYKyP+5sAZ3daPTGf7YMDh0sqw3b1hmxCvEOaLjOgD25w+CMlPsgM2y4TEm8NozrpERx7\u002F6yk106HYIqyXCFN5RqTAmL8VLBQZczeMiQvByIlgIicYPxeMnQlWnhzNoOZep7tSSzcIyw+M0VNBOM+YzAlGjgWhA0HPjrExuGfQlzXozRrMFhhS9WX5raGpkaUazgtmrgTTl4LhI8HQvqB\u002F1yB0qIO0ZjCwZxC5dg2DerqGAQtsqQVKGozVEmPxjq2mS\u002FeMtWdGvEyIPrpa7JWgtvvxKN2+sj9V06YboBuYqsvtJdW87mp6Vnyy6h7W2uY3s\u002F4I\u002F7GCtL7rPquNMmGt5Iz9BZQ5TSeButdkAAAAAElFTkSuQmCC","https:\u002F\u002Fcdn.sanity.io\u002Fimages\u002Fho74vmlz\u002Fproduction\u002F6b92362855c51347bc1e9531279166ee819079d7-2168x1080.png","teal","image","2026-04-23T16:31:28.000Z",[771,797,814],{"_id":772,"excerpt":773,"publishedAt":774,"slug":775,"thumbnail":776,"title":796},"25f54074-10d4-4f86-b855-6232676788e9","We're delighted to welcome Professor Susumu Kitagawa, winner of the 2025 Nobel Prize in Chemistry, as CuspAI’s latest advisor.","2026-11-19T10:00:00.000Z","nobel-laureate-susumu-kitagawa-joins-cuspais-advisory-board",{"alt":777,"asset":778,"crop":787,"hotspot":791},"Professor Susumu Kitagawa",{"_id":779,"metadata":780,"mimeType":349,"url":786},"image-d029c440c4e168db5800c4760636166bdfecbe12-3060x1722-png",{"dimensions":781,"lqip":785},{"aspectRatio":782,"height":783,"width":784},1.7770034843205575,1722,3060,"data:image\u002Fpng;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAALCAYAAAB\u002FCa1DAAAACXBIWXMAAAPoAAAD6AG1e1JrAAAC7klEQVQokVXS22+TdRzH8f47emcWjWHdwg41QYmpC8QLMHMLwQwSigGJikwEdkjECAluLDgXhoJZHTiVwE6Grufz0270fFiPtH3aPk\u002FbwZ2\u002Ft2nVqBeffO9e+eSTr6ZSkSmVSuTyWVLpGOFwEEly4vPZ8HqtuN0WXK4t\u002FAEnQZ8dk3EB4\u002FQ5Fic+Ym35LrlchrqqkMymkGIhoZHlGrJcpVwuUyjmSaXiPH0qEQi48fkdeL02XG4LTqcJv22T4G93sc5cwnjxJAvT53myuUYymyaeihKNh4WmWlOo1xVq9TrVapVCIUc8Huk03d7xEwh68PocWMzr2FcfEPz5Ns65KYwXxvjSMMLC\u002FE3sAQ9SRCIcCwpNrd5AUZsoavuqlMoldnfTJBJRotEQ4cg2wW0fDrsJ9\u002FoK0vK3WG9OcP\u002FSKWY\u002FPcH3i3P87jVjCznwRZxCU1caqI3W32lSkWVyuRzpdIpkMtGBI9EdpICbkGODyK+LOG9N8Wj6DPevfcba+hLWsBV7zIorahaadrt\u002FwRbVao1CoUA2myWTyZDZTZNs7xoJEvKZiT2+h3t+mpUJAz99M86WfQV\u002F2oEnYcUeMglNG2k092i2nndSr6s8K5UpFIrk8\u002FkOHE8k8AQktrY2cf1yB9PcZWbPvMfF0yPcvjeDRdrAHbFg2zELzX+x5t4L2hOUyjLF4jMSyTQ2h4eHj5\u002Fw44NVZm\u002Fd4frkZaY+PMbwW\u002F3oerQMjw7z1Y2rLD00smrd+D\u002FY2nuBojSoVKrk8kVMFjdXr3\u002FHJ+Nf8\u002FnkHIZzkwwdOspA3yCvv6qlq0vLvu4B3ta\u002Fy\u002FGTBsYnJv\u002Fa8B+0A6pN2r8ZT2X5YekRY4YvODJylrHTV3j\u002F+MfsH3yH1\u002Fbp6NMNMfDGIbQ9b9KtPUB3t46e3sH226hCbbREo7knmq3nQlGboiLXxHYoKWbml8Xo2AVx+MgpMfLBeXF09Kzo0w2J3n690B8+Jg7qh8X+fr3o6T0ourp6xcsvvfLHnw6JuYOIObvRAAAAAElFTkSuQmCC","https:\u002F\u002Fcdn.sanity.io\u002Fimages\u002Fho74vmlz\u002Fproduction\u002Fd029c440c4e168db5800c4760636166bdfecbe12-3060x1722.png",{"_type":788,"bottom":789,"left":790,"right":789,"top":789},"sanity.imageCrop",0,0.38344882933238,{"_type":792,"height":127,"width":793,"x":794,"y":795},"sanity.imageHotspot",0.61655117066762,0.6917244146661901,0.5,"Nobel Laureate Susumu Kitagawa joins CuspAI's advisory board",{"_id":798,"excerpt":799,"publishedAt":800,"slug":801,"thumbnail":802,"title":803},"post-medium-4f407afd45ea","Today, we’ve jointly published the largest open-access dataset to date for carbon capture materials in Nature Chemical Engineering. Working with Meta FAIR, Georgia Tech, and leading academic…","2026-10-29T16:06:17.000Z","opendac-2025-the-largest-open-dataset-for-direct-air-capture",{"alt":803,"asset":804,"crop":56,"hotspot":56},"OpenDAC 2025: the largest open dataset for direct air capture",{"_id":805,"metadata":806,"mimeType":812,"url":813},"image-63ee60b69350e431a64830e3bc72f421e0b25f98-2400x1348-jpg",{"dimensions":807,"lqip":811},{"aspectRatio":808,"height":809,"width":810},1.7804154302670623,1348,2400,"data:image\u002Fjpeg;base64,\u002F9j\u002F2wBDAAYEBQYFBAYGBQYHBwYIChAKCgkJChQODwwQFxQYGBcUFhYaHSUfGhsjHBYWICwgIyYnKSopGR8tMC0oMCUoKSj\u002F2wBDAQcHBwoIChMKChMoGhYaKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCj\u002FwAARCAALABQDASIAAhEBAxEB\u002F8QAGAAAAgMAAAAAAAAAAAAAAAAAAAMEBQf\u002FxAAhEAACAgEDBQEAAAAAAAAAAAABAgADERIhQQQFEzEycf\u002FEABYBAQEBAAAAAAAAAAAAAAAAAAMAAv\u002FEABkRAQEAAwEAAAAAAAAAAAAAAAEAERIhQf\u002FaAAwDAQACEQMRAD8Awbttmu5bWOWDb59kRnXUV1ZUEHnaVFLEAYPMa7t4mGo4gpnkSey3ZFbGISLb9D8hNaFX\u002F9k=","image\u002Fjpeg","https:\u002F\u002Fcdn.sanity.io\u002Fimages\u002Fho74vmlz\u002Fproduction\u002F63ee60b69350e431a64830e3bc72f421e0b25f98-2400x1348.jpg",{"_id":815,"excerpt":816,"publishedAt":817,"slug":818,"thumbnail":819,"title":827},"post-medium-f5603d8259dd","Today marks a milestone for the acceleration of materials discovery as we launch the ‘AI Materials Foundry’, a global network of data, labs, compute and scientific expertise for the design of new…","2026-07-20T06:04:59.000Z","launching-our-ai-materials-foundry-and-450-million-series-b",{"alt":820,"asset":821,"crop":56,"hotspot":56},"Launching our ‘AI Materials Foundry’ — and $450 million Series B",{"_id":822,"metadata":823,"mimeType":349,"url":826},"image-8c9dd3da26c30da6d4b83a57dbf5db2f5b76a715-2168x1080-png",{"dimensions":824,"lqip":825},{"aspectRatio":762,"height":763,"width":764},"data:image\u002Fpng;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAKCAYAAAC0VX7mAAAACXBIWXMAAAPoAAAD6AG1e1JrAAACvElEQVQokQXBfU8TBwDA4fsaS5aZuCyZcXMbiIDlxcooaaFUHVC0s7wEraNV7LFaWqHclTv6er3CcS0UoRZoAVeEbm6iIZroH2ZZ\u002FEy\u002FPY9wyeHhmz43X3QM8J3VztQ9L\u002FKcn3h4mshsAH9EwifnCagmStagqmucbRY52yuybSYJqybi0g7ibA7vyAzC1Ss2LN2D\u002FNRux3Ktj8C922gL06wtzZJWYoSVLLOZdeaN5+jGBgdaitOCxj\u002FPNCrFNKliBdV8gbJkEn4UQ7hw7luutVhxtNjot\u002FTweGqUouRjN+FnI7NIKpcnWdplpXpMubzLX4Us7zaznG5lqZdXOTh5yeHfpxweHVOr1hC+PHee1h9bcVzuwtXZzROfm634JDvxMYqJOXTDwKzW2Xt1xlG9zp\u002FrGV4XVBqmQmNL48ObQz5\u002FesXnjw3+e3+E0NzejNXahstuxet2IoemWJfuY0bvklPnyZUqGPsNysevqVb3qK0qvNAXqGWi7KUjHBgSjc04757L\u002FLufQBj71YX4yIsUnWZZCpJPzZNXIyQiD0knVPRnNZaKu4QSK0iKypaucLyR4o9ViXzEz8SQg9H+DmKTDioLYwi+8V94GnpAJv47emKONV3ByGsk4xKquoy6UiKo5PHMRPlNDFPIqJyUcrxcU0iH\u002FXRbWvjh4teMuzqRAkMIbmcboze68N2181T0oCyKyIvzhMQgD2eCBGNJxGWDGWWFOUWjpOc5KeqcFFLoUoje6120NX\u002FP+HAvAd8IgsfWxC3LRR78fAVpfIiof4IJtxOXrZNBp52x6SCP5SxPkmso2gb76xXelrc53TYxExJOex89nW14hgcYHr2FcOd6O97WJuSmq+j9wyxO3sHnbMZt+Yqb1kuMeDzcF6MEwjIxOUfd3OH9doW3myaFeIzBPhs93R3cvOGkt3+A\u002FwGo9ODJAa3VGwAAAABJRU5ErkJggg==","https:\u002F\u002Fcdn.sanity.io\u002Fimages\u002Fho74vmlz\u002Fproduction\u002F8c9dd3da26c30da6d4b83a57dbf5db2f5b76a715-2168x1080.png","Launching our ‘AI Materials Foundry’ — and $450 million Series B",{"metaDescription":56,"metaTitle":56,"noIndex":829,"ogImageUrl":56},false,"kups-a-molecular-simulation-engine-for-the-ai-era","kUPS: a molecular simulation engine for the AI era",[833,834,835,836,837,838,839,840,841],{"id":142,"text":139},{"id":241,"text":239},{"id":305,"text":303},{"id":396,"text":394},{"id":437,"text":435},{"id":541,"text":539},{"id":592,"text":590},{"id":627,"text":625},{"id":700,"text":698},[],1795092687760]