[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-121321-en":3,"doc-seo-121321-105":30,"detail-sidebar-cat-0-en-105":92},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":27,"seo_description":14,"update_tm":28,"read_time":29},121321,1099514067438,"River Wang","https://ap-avatar.wpscdn.com/avatar/100002539ee87300030?x-image-process=image/resize,m_fixed,w_180,h_180&k=1780474512215547542",8,"Research & Report","Reversible molecular simulation for training classical and machine-learning force fields - Research article","Next-generation force field development for molecular dynamics relies on abundant data, yet training with experimental information remains difficult, especially for machine-learning potentials. This work enhances differentiable molecular simulation by computing exact gradients via a reverse-time simulation, achieving effectively constant memory cost and a computation count comparable to forward runs. The method trains all-atom water and gas diffusion models and a machine-learning potential for diamond, matching time-dependent observables more accurately than ensemble reweighting.","RESEARCH ARTICLE  \nAPPLIED PHYSICAL SCIENCES  \n OPEN ACCESS  \nReversible molecular simulation for training classical and machine-learning force fields  \nJoe G. Greenera,1  \nEdited by Pablo Debenedetti, Princeton University, Princeton, NJ; received December 12, 2024; accepted April 22, 2025  \nThe next generation of force fields for molecular dynamics will be developed using a wealth of data. Training systematically with experimental data remains a challenge, however, especially for machine-learning potentials. Differentiable molecular simulation calculates gradients of observables with respect to parameters through molecular dynamics trajectories. Here, we improve this approach by explicitly calculating gradients using a reverse-time simulation with effectively constant memory cost anda computation count similar to the forward simulation. The method is applied to learn all-atom water and gas diffusion models with different functional forms and to train a machine-learning potential for diamond from scratch. Comparison to ensemble reweighting indicates that reversible simulation can provide more accurate gradientsand train to match time-dependent observables.  \ndifferentiable j reversible j molecular dynamics j force field  \nMolecular dynamics (MD) simulations have given us insight into how atoms move, from biomolecules to materials (1) . Key to the accuracy of a MD simulation is the accuracy of the force field used to describe how the atoms interact. For classical molecular mechanics, force field development has largely been manual with parameters tuned to give the best possible match to quantum mechanical (QM) data (bottom–up) and condensed phase properties (top–down) (2, 3). There have been automated approaches, including ensemble reweighting methods (4–8) like the popular ForceBalance (9–11), and graph neural networks to avoid discrete atom typing (12), but much work is still done manually (13) . The recently emerging and promising machine-learning interatomic potentials (MLIPs) (14, 15) are typically trained bottom–up on QM data alone (16), though this can give a distorted view of the utility of these models (17). While MLIPs can be validatedon other data (18), using non-QM data during training has proved challenging. This puts a heavy emphasis on generating large and diverse QM datasets and neglects other available data.  \nOne approach to training force fields with experimental data is differentiable molecular simulation (DMS), in which automatic differentiation (AD) (19) is used to obtain the gradients of a loss value with respect to the parameters over a simulation. This has had a number of recent applications (20–31) with dedicated software available (26, 32–35) . It is appealing due to the variety of possible loss functions and because the gradients are exact with respect to the forward simulation. There are, however, three main problems with DMS. First, the memory required is linear in the number of simulation steps meaning that gradient checkpointing is required for longer simulations (reducing the memory scaling to logarithmic (36)) and that larger neural networks maybe incompatible. Second, performance is considerably slower than standard simulation due to the overhead of reverse mode AD (RAD) . Finally, the gradients are prone to explosion due to the numerical integration. DMS holds promise despite this, particularly for training on time-dependent observables where ensemble reweighting approaches are not generally applicable (37) . Examples of these include diffusion coefficients, autocorrelation functions, relaxation rates, thermal conductivity, and reaction rates, where available data are challenging to use during training.  \nHere, we take inspiration from reversible differential equation solvers (38–40) and reversible neural networks (41–43) and ask whether DMS can be done without storing intermediate states, i.e., by explicitly deriving gradients rather than using conventional AD. This is motivated by three features o","cbCaidxrXO5ect2p","https://ap.wps.com/l/cbCaidxrXO5ect2p","pdf",4348481,1,11,"English","en",105,"# Introduction\n## Key challenges in force-field training\n# Reversible differentiable molecular simulation\n## Motivation and gradient computation strategy\n# Results and demonstrations\n## Water and gas diffusion models\n## Machine-learning potential for diamond\n## Comparison to ensemble reweighting","[{\"question\":\"What problem does reversible molecular simulation address in force-field training?\",\"answer\":\"It targets the difficulty of training classical and machine-learning force fields using experimental or time-dependent observables by improving gradient computation during simulation-based optimization.\"},{\"question\":\"How does the proposed method compute gradients more efficiently than standard differentiable molecular simulation?\",\"answer\":\"It explicitly calculates gradients using a reverse-time simulation, yielding effectively constant memory cost while keeping computation comparable to forward simulation.\"},{\"question\":\"Which kinds of models and observables are used to demonstrate the approach?\",\"answer\":\"The method is applied to learn all-atom water and gas diffusion models and to train a machine-learning potential for diamond, with performance assessed on time-dependent observables.\"}]","Reversible molecular simulation for training classical and machine-learning force fields - Research article | PDF",1785735062,28,{"code":4,"msg":31,"data":32},"ok",{"site_id":24,"language":23,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":87,"head_meta":89,"extra_data":91,"updated_unix":28},"reversible-molecular-simulation-for-training-classical-and-machine-learning-force-fields-research-article","",{"@graph":36,"@context":86},[37,54,69],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,51],{"item":41,"name":42,"@type":43,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":50},"https://docshare.wps.com/document/research-report/",3,{"item":52,"name":13,"@type":43,"position":53},"https://docshare.wps.com/document/reversible-molecular-simulation-for-training-classical-and-machine-learning-force-fields-research-article/121321/",4,{"url":52,"name":13,"@type":55,"author":56,"headline":13,"publisher":58,"fileFormat":61,"inLanguage":23,"description":14,"dateModified":62,"datePublished":63,"encodingFormat":61,"isAccessibleForFree":64,"interactionStatistic":65},"DigitalDocument",{"name":9,"@type":57},"Person",{"url":41,"name":59,"@type":60},"DocShare","Organization","application/pdf","2026-08-04","2026-08-03",true,{"@type":66,"interactionType":67,"userInteractionCount":20},"InteractionCounter",{"@type":68},"ViewAction",{"@type":70,"mainEntity":71},"FAQPage",[72,78,82],{"name":73,"@type":74,"acceptedAnswer":75},"What problem does reversible molecular simulation address in force-field training?","Question",{"text":76,"@type":77},"It targets the difficulty of training classical and machine-learning force fields using experimental or time-dependent observables by improving gradient computation during simulation-based optimization.","Answer",{"name":79,"@type":74,"acceptedAnswer":80},"How does the proposed method compute gradients more efficiently than standard differentiable molecular simulation?",{"text":81,"@type":77},"It explicitly calculates gradients using a reverse-time simulation, yielding effectively constant memory cost while keeping computation comparable to forward simulation.",{"name":83,"@type":74,"acceptedAnswer":84},"Which kinds of models and observables are used to demonstrate the approach?",{"text":85,"@type":77},"The method is applied to learn all-atom water and gas diffusion models and to train a machine-learning potential for diamond, with performance assessed on time-dependent observables.","https://schema.org",{"og:url":52,"og:type":88,"og:title":13,"og:site_name":59,"og:description":14},"article",{"robots":90,"canonical":52},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":93},[94,98,102,106,111,116,121,124,129,132,136],{"id":20,"doc_module":4,"doc_module_name":46,"category_name":95,"show_sort_weight":96,"slug":97},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":99,"show_sort_weight":100,"slug":101},"Literature",80,"literature",{"id":53,"doc_module":4,"doc_module_name":46,"category_name":103,"show_sort_weight":104,"slug":105},"Exam",70,"exam",{"id":107,"doc_module":4,"doc_module_name":46,"category_name":108,"show_sort_weight":109,"slug":110},5,"Comic",60,"comic",{"id":112,"doc_module":4,"doc_module_name":46,"category_name":113,"show_sort_weight":114,"slug":115},6,"Technology",50,"technology",{"id":117,"doc_module":4,"doc_module_name":46,"category_name":118,"show_sort_weight":119,"slug":120},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":122,"slug":123},30,"research-report",{"id":125,"doc_module":4,"doc_module_name":46,"category_name":126,"show_sort_weight":127,"slug":128},9,"Religion & Spirituality",20,"religion-spirituality",{"id":127,"doc_module":4,"doc_module_name":46,"category_name":130,"show_sort_weight":127,"slug":131},"World Cup","world-cup",{"id":133,"doc_module":4,"doc_module_name":46,"category_name":134,"show_sort_weight":133,"slug":135},10,"Lifestyle","lifestyle",{"id":137,"doc_module":4,"doc_module_name":46,"category_name":138,"show_sort_weight":107,"slug":139},19,"General","general"]