[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-82007-en":3,"doc-seo-82007-105":29,"detail-sidebar-cat-0-en-105":91},{"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":13,"seo_description":14,"update_tm":27,"read_time":28},82007,687197207639,"Asher","https://ap-avatar.wpscdn.com/davatar_a8503ba1806abce46bf441b54a3ca4cd",8,"Research & Report","Are Machine Learning Interatomic Potentials Truly Practical A Benchmark of 23 Mainstream Models","Most MLIP benchmarks emphasize static predictive accuracy while neglecting inference efficiency and hardware scalability, encouraging model bloat with limited practical impact. This study benchmarks 23 mainstream open-source MLIPs using a unified ASE-based pipeline on a low-cost NVIDIA DGX Spark platform, with memory capped at 80 GB and a fixed 192-atom system. Three metrics—predictive accuracy, MD throughput, and atomic scalability—reveal a strong accuracy–efficiency trade-off and show lightweight MLIPs better align with the Pareto frontier.","Are Machine Learning Interatomic Potentials Truly Practical?  \nA Benchmark of 23 Mainstream Models  \nHanwen Kang 1,2, Tenglong Lu 1,3, Sheng Meng 1 *, Miao Liu 1,3 *  \n1Institute of Physics, Chinese Academy of Sciences, Beijing 100190, China  \n2School of Physical Sciences, University of Chinese Academy of Sciences, Beijing 100049, China 3Dongguan Institute of Materials Science and Technology, Dongguan, Guangdong 523808, China  \n*Corresponding author: [smeng@iphy.ac.cn](smeng@iphy.ac.cn), [mliu@iphy.ac.cn](mliu@iphy.ac.cn)  \nAbstract. Most MLIP benchmarks reward static accuracy while ignoring inference efficiency and hardware scalability — driving model bloat with unclear real-world value. We benchmark 23 mainstream open-source MLIPs on a low-cost NVIDIA DGX Spark (128 GB native memory, capped at 80 GB to mimic ordinary lab hardware), using a fixed 192-atom system under a unified ASE-based pipeline. We evaluate three dimensions: predictive accuracy, MD simulation throughput, and atomic scalability. Our results expose a sharp accuracy–efficiency trade-off: large SOTA models deliver only 3–5 meV/atom more accuracy than lightweight ones, but lose orders of magnitude in throughput—in the worst case, becoming only marginally faster than DFT itself. Lightweight MLIPs, by contrast, sit on the Pareto frontier and run on modest hardware. The lesson is that single-dimensional benchmarks mislead the field, and that future MLIP development should value efficiency and scalability alongside accuracy.  \n1. Introduction  \nAtomistic molecular dynamics (MD) is how materials science has, for half a century, asked what atoms actually do at room temperature. The catch has always been the choice between two inadequate tools.  \nDensity functional theory (DFT) [1–3] delivers quantum-level accuracy but scales unfavorably—its cost grows roughly with the cube of the number of atoms (O(N3)) . A few  \nhundred atoms is comfortable; a few thousand is a research quarter. Classical empirical force fields scale linearly and are blazingly fast, but they cannot see the subtle environment-dependent bonding that defines much of materials chemistry. For decades, the field has lived with this trade-off.  \nMachine-learning interatomic potentials (MLIPs) were supposed to dissolve it. By encoding local atomic environments within a 6–7 Å cutoff and learning high-dimensional potentialenergy surfaces from massive DFT datasets, modern MLIPs approach DFT accuracy (energy MAEs of 20–30 meV/atom; force errors in the tens of meV·Å⁻¹) while inheriting force-fieldlike linear scaling [4][5] . A diverse open-source ecosystem—Equiformer [6], DPA [7–9], PET [10,11], M3GNet [12], CHGNet [13], MatterSim [14], GRACE [15,16], MACE [17,18], GPTFF [19], MatRIS [20], NequIP [21–23], TACE [24], EquFlash, eSEN [25],SevenNet [26– 28], ORB [29,30], Nequix [31,32], Allegro [33] — has proliferated, each nudging the Matbench Discovery leaderboard a little higher.  \nAnd then something went quietly wrong.  \nWhile accuracy improved, model size exploded. State-of-the-art universal MLIPs now exceed 730 million parameters [11]—a scale that dwarfs the size of individual training datasets. The accuracy gains from this expansion are real but small: roughly 3–5 meV/atom, smaller than room-temperature thermal noise. Meanwhile, inference throughput drops, memory consumption rises, and the maximum size of simulatable systems collapses. In the most extreme cases, large MLIPs run less than 2x faster than DFT, surrendering the central efficiency advantage that justified their existence.  \nThis paper is a systematic benchmark study designed to expose that trade-off. We test 23 mainstream open-source MLIPs on standardized hardware and ask three questions: how accurate are they really? how fast? and how big a system can they actually simulate? Our goal is not to crown a winner—it is to provide a clear, multi-dimensional view of the current MLIP landscape, and to give working researchers a practical basis for choosin","cbCaifiClsK8Gqzd","https://ap.wps.com/l/cbCaifiClsK8Gqzd","pdf",2014741,1,13,"English","en",105,"# Abstract\n# 1. Introduction\n# 2. Methods","[{\"question\":\"Why do current MLIP benchmarks fail to reflect real-world practicality?\",\"answer\":\"They often reward static accuracy but ignore inference efficiency and hardware scalability, which allows model size to grow without clear practical value.\"},{\"question\":\"How was the benchmark set up to ensure comparability across models?\",\"answer\":\"The study evaluates 23 mainstream open-source MLIPs on standardized NVIDIA DGX Spark hardware using a unified ASE-based pipeline, with cross-validation on TorchSim for bias control and an 80 GB memory cap.\"},{\"question\":\"What key trade-off do the results reveal about MLIP model size?\",\"answer\":\"Large state-of-the-art MLIPs provide only a small accuracy gain (about 3–5 meV/atom) while losing orders of magnitude in MD throughput, sometimes becoming only marginally faster than DFT.\"}]",1784177540,33,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":27},"are-machine-learning-interatomic-potentials-truly-practical-a-benchmark-of-23-mainstream-models","",{"@graph":35,"@context":85},[36,53,68],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/are-machine-learning-interatomic-potentials-truly-practical-a-benchmark-of-23-mainstream-models/82007/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-17","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"Why do current MLIP benchmarks fail to reflect real-world practicality?","Question",{"text":75,"@type":76},"They often reward static accuracy but ignore inference efficiency and hardware scalability, which allows model size to grow without clear practical value.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How was the benchmark set up to ensure comparability across models?",{"text":80,"@type":76},"The study evaluates 23 mainstream open-source MLIPs on standardized NVIDIA DGX Spark hardware using a unified ASE-based pipeline, with cross-validation on TorchSim for bias control and an 80 GB memory cap.",{"name":82,"@type":73,"acceptedAnswer":83},"What key trade-off do the results reveal about MLIP model size?",{"text":84,"@type":76},"Large state-of-the-art MLIPs provide only a small accuracy gain (about 3–5 meV/atom) while losing orders of magnitude in MD throughput, sometimes becoming only marginally faster than DFT.","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":45,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":45,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":45,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":45,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":45,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":45,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":45,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]