rationale:

Overall Outlook and Bottlenecks While general-purpose AI systems are advancing rapidly—as seen in mathematics, where performance on challenging benchmarks leaped to 100% and major conjectures were resolved by 2026—translating this to physical materials faces severe bottlenecks. Progress in materials science requires physical synthesis, scale-up, and lengthy validation. Published closed-loop synthesis throughput grew only modestly between 2023 and 2026, reaching about 32 samples a day. Furthermore, the criteria for this assessment are strict: the material must be first proposed by a general-purpose AI (specialist ML models alone do not count), and it must hold the peer-reviewed published record on exactly 2030-12-31. Peer review alone introduces a 1–2 year lag, meaning discoveries made after mid-2029 are unlikely to qualify. At the same time, human research groups are continuously raising the benchmarks across all these fields.

Most Plausible Categories: Catalysts and Sorbents Iridium-free acidic water-splitting catalysts and humid direct-air-capture (DAC) sorbents represent the most promising avenues for AI records. In the catalyst space, automated platforms have already screened thousands of candidates, with one lead candidate (InMnPdOx) maintaining low overpotential for 1,000 hours Lila Sciences' AI Lab Uncovers Palladium Catalysts for ... https://arxiv.org/pdf/2609.30133. However, this test was run at a low current density (10 mA/cm²), and the result remains a preprint. The human-established record includes a fluorinated RuO₂ anode running for 1,440 hours at up to 1 A/cm² Durable acidic water oxidation ruthenium based ... Durable acidic water oxidation ruthenium based ..., and durability must be compared at fixed activity, making it a highly contestable record. For DAC sorbents, human-developed materials like COF-999 (2.05 mmol/g) Carbon dioxide capture from open air using covalent ... and COF-1000 (2.19 mmol/g) Fast and selective CO2 capture from outdoor air by covalent organic ... set a high, moving bar. While agent-guided workflows have begun synthesizing new COFs, demonstrating record cyclic working capacity under humid conditions before the deadline remains a significant hurdle.

Less Plausible Categories: Conductors, Superconductors, and Magnets Solid lithium-ion conductors face a daunting human-held benchmark of 42 mS/cm Scandium Induced Structural Disorder and Vacancy .... Recent general AI lab campaigns achieved conductivities around 0.14 mS/cm Agentic LLM Reasoning in a Self-Driving Laboratory for Air ..., suggesting that while AI screening is viable, reaching a verified peer-reviewed record by 2030 is a long shot. Ambient-pressure superconductors and rare-earth-free permanent magnets are the least likely to fall to AI by 2030. Identifying new high-Tc families is notoriously difficult; human-made nickelate films have reached a 63 K onset, whereas the best experimentally validated AI-proposed superconductor remains stalled at 5.4 K. Similarly, permanent magnet energy product depends heavily on microstructure and processing—not just chemical composition—making autonomous optimization significantly harder.

Correlations and Final Distribution Because these five categories share the same fundamental physical and institutional bottlenecks, their outcomes are strongly correlated. If autonomous labs and agentic science manage to scale 10–100× faster than currently demonstrated, it could trigger a fast-takeoff scenario where multiple material categories flip simultaneously. Set against related questions, the timelines for physical validation support a moderate number of breakthroughs by 2030: there is time for some discoveries to clear synthesis and peer review, but not enough to assume broad success across all five categories. I therefore place the median at 1.5 records, between one and two qualifying records, with the 25th and 75th percentiles at 0.6 and 2.5. These fractional values describe the forecast's percentile positions, not fractional material records; the realized count must be a non-negative whole number. The 10th percentile is zero, reflecting the possibility that attribution, validation, or moving human benchmarks prevent any qualifying record. The positively correlated upper tail puts the 90th percentile at 3.5, allowing several categories to succeed together if automation scales rapidly. This distribution balances the physical friction limiting the lower outcomes with a moderate central outlook and meaningful upside from a broader automation breakthrough.