{"id":55090,"date":"2026-07-20T11:50:05","date_gmt":"2026-07-20T15:50:05","guid":{"rendered":"https:\/\/engineering.jhu.edu\/materials\/?post_type=news&p=55090"},"modified":"2026-07-23T12:36:18","modified_gmt":"2026-07-23T16:36:18","slug":"teaching-ai-to-design-new-materials-by-working-backwards","status":"publish","type":"news","link":"https:\/\/engineering.jhu.edu\/materials\/news\/teaching-ai-to-design-new-materials-by-working-backwards\/","title":{"rendered":"Teaching AI to Design New Materials By Working Backwards"},"content":{"rendered":"

Imagine telling a computer, “I want a material that\u00a0is hard, durable, and\u00a0conducts electricity at room temperature\u00a0without giving off extra heat,” and having it hand back\u00a0an\u00a0idea\u00a0to test in the lab. That’s the goal of a growing area of AI research called\u00a0\u201cinverse materials design,\u201d\u00a0a field of study where researchers\u00a0give\u00a0AI algorithms with prompts on\u00a0the qualities and functions they want from a material, and the\u00a0AI tool\u00a0works backward\u00a0to\u00a0provide\u00a0options\u00a0the scientists can test in the laboratory.<\/span><\/p>\n

Several\u00a0companies and research labs have\u00a0created models to provide such reverse-engineering guidance, but\u00a0a team\u00a0from\u00a0Johns Hopkins and West Virginia University\u00a0says\u00a0it’s\u00a0hard to know what\u00a0aspects of the models works\u00a0well.\u00a0<\/span>\u00a0<\/span><\/p>\n

To\u00a0compare the models, the research team created <\/span>AtomBench <\/span><\/i>to explore the models\u2019 methods and rank their ability to inform scientists. <\/span>\u00a0<\/span><\/p>\n

A description of\u00a0<\/span>AtomBench<\/span><\/i><\/a>\u00a0is published\u00a0this month\u00a0in\u00a0the Royal Society of Chemistry\u2019s\u00a0journal\u00a0<\/span>Digital Discovery.<\/span><\/i>\u00a0<\/span><\/p>\n

Until now, materials science has\u00a0worked in one direction: scientists study a material’s atomic structure, then use experiments or computers to investigate the substance\u2019s\u00a0properties\u00a0–\u00a0hardness, conductivity, magnetism. This\u00a0process\u00a0has\u00a0enabled\u00a0materials scientists to create stronger metals, better batteries,\u00a0and\u00a0safer medical implants, but the process of material design can be slow and starts with materials that already exist. With inverse materials\u00a0design, researchers\u00a0may be able to\u00a0use AI models to\u00a0reverse-engineer\u00a0that process,\u00a0going\u00a0from property to structure.\u00a0<\/span>\u00a0<\/span><\/p>\n

\u201cIn recent years, it feels like AI has gotten its PhD in materials science,\u201d says\u00a0<\/span>Kamal\u00a0Choudhary<\/span><\/a>, an\u00a0assistant\u00a0professor in the departments of Materials Science and Engineering and Electrical Engineering and Computer Science at Johns Hopkins and\u00a0leader of the research.\u00a0\u201cAI models could soon reach a Nobel laureate level of science. Our goal is to\u00a0help\u00a0every lab\u00a0feel like they have a Nobel laureate at their fingertips.\u201d<\/span>\u00a0<\/span><\/p>\n

For the current study,\u00a0the team focused\u00a0on\u00a0superconductors\u00a0\u2014 materials that, below a certain temperature, conduct electricity without the resistance and heat that\u00a0most materials\u00a0emit\u00a0when conducting electricity.\u00a0<\/span>\u00a0<\/span><\/p>\n

Choudhary says that finding a superconductor that works at room temperature\u00a0could\u00a0transform the sustainability, efficiency, and cost of power grids and electronics.\u00a0Finding one, though, means searching an\u00a0infinite\u00a0number\u00a0of\u00a0potential\u00a0materials.\u00a0<\/span>\u00a0<\/span><\/p>\n

Choudhary and the team\u00a0tested\u00a0four AI models for materials design <\/span>\u2014 <\/span>CDVAE<\/span><\/i>,\u00a0developed by MIT researchers, <\/span>FlowMM <\/span><\/i>from MetaAI and the University of Amsterdam,\u00a0<\/span>MatterGen<\/span><\/i>\u00a0from Microsoft\u00a0Research AI, and one they developed,\u00a0<\/span>AtomGPT<\/span><\/i><\/a>. Each\u00a0model\u00a0uses\u00a0different approaches to solving materials design problems and\u00a0analyzing data.\u00a0<\/span>\u00a0<\/span><\/p>\n

The team then gave the\u00a0models\u00a0the same task: take\u00a0the\u00a0chemical composition\u00a0for a superconductor\u00a0and\u00a0predict\u00a0the\u00a0spatial\u00a0positions of its atoms.\u00a0<\/span>\u00a0<\/span><\/p>\n

The models\u00a0were grouped\u00a0by\u00a0the number of prompts needed to level the playing field\u00a0before making their\u00a0predictions.\u00a0\u00a0<\/span>\u00a0<\/span><\/p>\n

The results\u00a0did not\u00a0identify\u00a0a clear winner.\u00a0<\/span>\u00a0<\/span><\/p>\n

Microsoft’s\u00a0<\/span>MatterGen<\/span><\/i>\u00a0was the most\u00a0accurate\u00a0at placing individual atoms correctly, with Choudhary\u2019s\u00a0<\/span>AtomGPT<\/span><\/i>\u00a0a close second.\u00a0<\/span>CDVAE\u00a0<\/span><\/i>was the best at reproducing the overall shape and\u00a0size of the structure, and Meta\u2019s <\/span>FlowMM <\/span><\/i>was the fastest but trailed the others in accuracy.\u00a0<\/span>\u00a0<\/span><\/p>\n

Choudhary says one of the results that surprised him was that <\/span>MatterGen <\/span><\/i>achieved its\u00a0prediction\u00a0in atomic accuracy\u00a0with\u00a0about\u00a0140 times fewer\u00a0parameters than the ones researchers gave <\/span>AtomGPT<\/span><\/i>. The researchers attribute this to\u00a0<\/span>MatterGen<\/span><\/i>‘s\u00a0training: it was trained on large databases of real\u00a0molecular\u00a0structures and built with physical constraints baked into its architecture.\u00a0<\/span>\u00a0<\/span><\/p>\n

AtomGPT<\/span><\/i>\u00a0uses\u00a0a\u00a0model\u00a0similar to\u00a0popular\u00a0AI models <\/span>ChatGPT <\/span><\/i>and <\/span>Claude <\/span><\/i>which employ\u00a0a\u00a0large\u00a0language\u00a0model algorithm\u00a0to process human language.\u00a0For Choudhary, what this means is that\u00a0a smaller model trained with the\u00a0<\/span>right<\/span><\/i>\u00a0knowledge and physical “common sense” constraints, like <\/span>MatterGen, <\/span><\/i>can outperform a much larger general-purpose model trained on\u00a0more data, which he believes is\u00a0a finding with implications beyond materials science.<\/span>\u00a0<\/span><\/p>\n

The team\u2019s goal, they say,\u00a0is to make\u00a0the comparative\u00a0information useful to materials scientists.\u00a0<\/span>\u00a0<\/span><\/p>\n

The project\u2019s first author, Rhys Campbell,\u00a0began this work as an undergraduate intern in Choudhary’s group\u00a0at NIST.\u00a0Campbell will join Choudhary\u2019s lab at Hopkins as a PhD student this fall.<\/span>\u00a0<\/span><\/p>\n

\u201cCreating a new statistical metric to evaluate these models and using information theory to determine if the models are being fairly compared was my favorite part, and empirically validating these theoretical developments was quite satisfying,\u201d says Campbell, who recently received his\u00a0bachelor’s degree in physics\u00a0from West Virginia University. \u201cI can’t thank my mentors, Kamal Choudhary and Aldo Romero, enough for guiding me through this process and teaching me how to think like a researcher.\u201d<\/span>\u00a0<\/span><\/p>\n

The\u00a0<\/span>open-source<\/span><\/a>\u00a0<\/span>AtomBench<\/span><\/i>\u00a0tool is available to scientists. Choudhary is continuing to develop\u00a0AtomGPT\u00a0with the goal of making it\u00a0\u201ca ChatGPT for superconductors\u201d\u00a0and potentially other materials.\u00a0Users can enter a chemical formula and a target temperature at which\u00a0they\u2019d\u00a0like their superconductor to\u00a0work\u00a0and the tool provides\u00a0a\u00a0possible structure.<\/span>\u00a0<\/span><\/p>\n","protected":false},"template":"","class_list":["post-55090","news","type-news","status-publish","hentry","news_categories-research"],"acf":[],"yoast_head":"\nTeaching AI to Design New Materials By Working Backwards - Department of Materials Science & Engineering<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/engineering.jhu.edu\/materials\/news\/teaching-ai-to-design-new-materials-by-working-backwards\/\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Teaching AI to Design New Materials By Working Backwards - Department of Materials Science & Engineering\" \/>\n<meta property=\"og:description\" content=\"Imagine telling a computer, “I want a material that\u00a0is hard, durable, and\u00a0conducts electricity at room temperature\u00a0without giving off extra heat,” and having it hand back\u00a0an\u00a0idea\u00a0to test in the lab. That’s…\" \/>\n<meta property=\"og:url\" content=\"https:\/\/engineering.jhu.edu\/materials\/news\/teaching-ai-to-design-new-materials-by-working-backwards\/\" \/>\n<meta property=\"og:site_name\" content=\"Department of Materials Science & Engineering\" \/>\n<meta property=\"article:modified_time\" content=\"2026-07-23T16:36:18+00:00\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data1\" content=\"4 minutes\" \/>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Teaching AI to Design New Materials By Working Backwards - Department of Materials Science & Engineering","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/engineering.jhu.edu\/materials\/news\/teaching-ai-to-design-new-materials-by-working-backwards\/","og_locale":"en_US","og_type":"article","og_title":"Teaching AI to Design New Materials By Working Backwards - Department of Materials Science & Engineering","og_description":"Imagine telling a computer, “I want a material that\u00a0is hard, durable, and\u00a0conducts electricity at room temperature\u00a0without giving off extra heat,” and having it hand back\u00a0an\u00a0idea\u00a0to test in the lab. That’s…","og_url":"https:\/\/engineering.jhu.edu\/materials\/news\/teaching-ai-to-design-new-materials-by-working-backwards\/","og_site_name":"Department of Materials Science & Engineering","article_modified_time":"2026-07-23T16:36:18+00:00","twitter_card":"summary_large_image","twitter_misc":{"Est. reading time":"4 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/engineering.jhu.edu\/materials\/news\/teaching-ai-to-design-new-materials-by-working-backwards\/","url":"https:\/\/engineering.jhu.edu\/materials\/news\/teaching-ai-to-design-new-materials-by-working-backwards\/","name":"Teaching AI to Design New Materials By Working Backwards - Department of Materials Science & Engineering","isPartOf":{"@id":"https:\/\/engineering.jhu.edu\/materials\/#website"},"datePublished":"2026-07-20T15:50:05+00:00","dateModified":"2026-07-23T16:36:18+00:00","breadcrumb":{"@id":"https:\/\/engineering.jhu.edu\/materials\/news\/teaching-ai-to-design-new-materials-by-working-backwards\/#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/engineering.jhu.edu\/materials\/news\/teaching-ai-to-design-new-materials-by-working-backwards\/"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/engineering.jhu.edu\/materials\/news\/teaching-ai-to-design-new-materials-by-working-backwards\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/engineering.jhu.edu\/materials\/"},{"@type":"ListItem","position":2,"name":"News","item":"https:\/\/engineering.jhu.edu\/materials\/news\/"},{"@type":"ListItem","position":3,"name":"Teaching AI to Design New Materials By Working Backwards"}]},{"@type":"WebSite","@id":"https:\/\/engineering.jhu.edu\/materials\/#website","url":"https:\/\/engineering.jhu.edu\/materials\/","name":"Department of Materials Science & Engineering","description":"Department of Materials Science & Engineering","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/engineering.jhu.edu\/materials\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"}]}},"distributor_meta":false,"distributor_terms":false,"distributor_media":false,"distributor_original_site_name":"Department of Materials Science & Engineering","distributor_original_site_url":"https:\/\/engineering.jhu.edu\/materials","push-errors":false,"_links":{"self":[{"href":"https:\/\/engineering.jhu.edu\/materials\/wp-json\/wp\/v2\/news\/55090","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/engineering.jhu.edu\/materials\/wp-json\/wp\/v2\/news"}],"about":[{"href":"https:\/\/engineering.jhu.edu\/materials\/wp-json\/wp\/v2\/types\/news"}],"wp:attachment":[{"href":"https:\/\/engineering.jhu.edu\/materials\/wp-json\/wp\/v2\/media?parent=55090"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}