Three engineers with pedigrees from Stanford, MIT, and Harvard Business School believe the answer to the next battery breakthrough or PFAS replacement sits in forgotten lab notebooks, not supercomputers. Their three-person startup, 83 Sciences, says it has already discovered a novel material by training AI models on unpublished experimental data from research labs: the 85 percent of results that typically get discarded. The San Francisco-based team, backed by Y Combinator, claims that lost data represents more than $100 billion in annual R&D value and promises to turn a partner's discarded experiments into a co-authored paper in under two months.
The pitch arrives at a moment when regulatory deadlines and semiconductor industrial policy have created urgent demand for breakthrough materials. The EPA finalized the first national drinking water standards for six PFAS compounds in April 2024, according to agency documents. EU restrictions on PFAS in food-contact packaging took effect on August 11, 2026. The CHIPS Act has committed $11 billion to R&D programs, including funding commitments announced by NIST. Those pressures are colliding with a widening gap between computational materials databases and high-throughput experimental datasets. The Materials Project holds more than 130,000 compounds; AFLOW has roughly 3.55 million. NREL's experimental HTEM dataset, by contrast, contains around 80,000 samples, a National Science Review study noted in March 2026.
Mining the file drawer
Google DeepMind's GNoME model identified 2.2 million candidate crystal structures in November 2023, with roughly 380,000 predicted stable, equivalent to about 800 years' worth of knowledge by the company's count. Meta FAIR released its UMA family of universal atomic models in May 2025, followed by Open Molecules 2025 and OpenDAC datasets. Microsoft and Pacific Northwest National Laboratory narrowed 32 million electrolyte candidates to 23 options using cloud AI and high-performance computing, then synthesized a new Na_xLi_{3-x}YCl_6 solid-electrolyte family, according to an arXiv preprint published in January 2024. Berkeley's A-Lab runs a closed-loop synthesis and characterization system targeting 10-to-100-times speedups. Rice University received a $19.9 million NSF award in July 2026 for an AI-powered autonomous materials lab called READINESS. The system will learn from both successful and unsuccessful experiments, the university said.
Most of those advances rely on published data or computational screening. Yet perhaps the richest vein of information sits in what researchers discard. "The vast majority of data generated in research labs never gets published or used again," Yankang Yang, one of 83 Sciences' three cofounders, wrote on LinkedIn roughly a month before the company's Y Combinator launch. A 2016 Nature paper on the Dark Reactions Project demonstrated the value of mining failed hydrothermal syntheses from lab notebooks. A Royal Society of Chemistry review published in January 2026 warned that materials datasets "overrepresent stable equilibrium-phase systems," risking biased AI models.
The company's website lists five priority verticals: energy storage, critical minerals and metals, catalysis and chemicals, life sciences solid forms, and semiconductors. Its Y Combinator launch cited PFAS replacements and silicon-anode binder discovery as examples, referencing CEO Ian Naccarella's prior work at Sila Nanotechnologies. A gated early-access page for the Dalton Electronic Lab Notebook and Data Structuring Platform was live as of late August 2026. The site says 83 Sciences mines "file-drawer" data from lab notebooks, instrument files, and voice notes to build "a structured, queryable agentic-record of every experiment" that "understands failures and proposes optimized process conditions." Engagement models include industry discovery contracts and academic partnerships producing papers, patents, and commercialization, according to the company.
Credentials and competition
Naccarella holds a Stanford BS and MS in chemical engineering and an HBS MBA; he previously worked in strategy at Sila, a silicon-anode battery startup, and consulted at BCG. Cofounder and CTO Eric Riesel earned a PhD in inorganic chemistry from MIT, where he researched "one of the first generative approaches for chemical experimental data," the company's Y Combinator profile states. Yang, the COO, led BCG's AI program to more than 30,000 users and has a Harvard CS and electrical engineering background. The startup is hiring a founding AI engineer at $120,000 to $250,000, a head of data partnerships at $170,000 to $220,000, and a full-stack engineer at $160,000 to $180,000, job postings on its Y Combinator page showed.
The startup has not disclosed funding beyond Y Combinator, customers, or a public preprint of its claimed novel material discovery. Its competitor landscape includes Citrine Informatics, which raised a $16 million Series C announced in January 2023 and has published case studies with Huntsman Building Solutions on polyurethane foam formulation and PFAS alternatives, according to an AZoMaterials interview dated April 22, 2026. CuspAI, cofounded by machine learning researcher Max Welling, raised a $30 million seed in June 2024 from Hoxton Ventures, Basis Set, Lightspeed, and FJ Labs to build an "AI Materials Foundry" with NVIDIA and Meta FAIR, SiliconANGLE reported. Orbital Materials announced a Series A of $16 million in 2024, with secondary reports claiming a $50 million Series B in September 2026. Another Y Combinator company, Discovered Materials, unveiled a "Material Discovery Bench" for semiconductor materials in a TechCrunch profile on August 10, 2026, acknowledging the need for wet-lab validation. Kebotix, Intellegens, and Aionics also offer AI-driven materials platforms.
IDTechEx projected materials informatics provider revenues above $700 million by 2034, with a compound annual growth rate of 13.7 percent through 2033, according to reports updated in 2023 and 2024. An ACS Omega paper published in July 2026 cautioned against "scaling AI without preserving or expanding experimental infrastructure and provenance." An arXiv preprint from August 2026 titled "When Literature Data Mislead AI in Materials Discovery" warned of label noise and biases in literature-derived datasets focused on solid electrolytes. "Transparency is crucial for understanding and correcting biases in AI models," an OECD report on AI in science stated in 2023.
Signals in the noise
Rice's engineering dean, Luay Nakhleh, said in the July 2026 READINESS announcement that "responsible AI should complement researchers' capabilities rather than replace their judgment." David Sholl, Rice's executive vice president for research, added that "by lowering those barriers, READINESS can accelerate discovery and expand who can participate." The NSF program and similar consortia signal a shift toward explicitly capturing negative results, something the academic publishing system has long discouraged. DeepMind called for data "stocktakes" and access to "non-traditional" sources like researchers' logbooks in a 2025 policy note, warning that AI can overfit anomalies without raw experimental context.
The vendor ecosystem is bifurcating. Foundation model and open-dataset efforts from DeepMind, Meta FAIR, and others provide a computational layer that many platforms now build on. Domain-specialist vendors integrating proprietary experimental data, electronic lab notebooks, and self-driving lab workflows represent a second tier. 83 Sciences positions itself in that second camp, betting that the failures sitting in decades of lab notebooks hold signals computational models miss.
Whether the company can execute on its two-month manuscript timeline and scale partnerships across energy storage, PFAS alternatives, and semiconductors will test the value proposition. The founders' prior experience at Sila, BCG's AI program, and MIT's generative-chemistry research gives them credibility, though no external funding or customer names limit third-party validation for now. The company is accepting inquiries at [email protected], its Y Combinator launch page noted. If the team is right, the next materials revolution may not require a supercomputer, just a willingness to look where others have already stopped looking.
