rationale:

AI Design with Human Execution (33%) The most likely path to a Nobel-caliber materials breakthrough is an AI system designing a material or experiment that human scientists then synthesize and validate. This route leverages the world's massive existing laboratory infrastructure and replication networks without waiting for the formidable engineering challenges of end-to-end automation to be fully solved. Physical scale-up bottlenecks pose a particularly severe hurdle for fully autonomous discovery workflows in materials science, further favoring designs that human scientists can execute using existing infrastructure. Current cutting-edge AI materials work already fits this paradigm: a GPT-based agent designed the synthesis conditions for COF-2000 New AI agent is 'a paradigm shift' for COF synthesis, Gemini provided recipes for semi-automated CVD reactors with human handling Accelerating Scientific Research with Gemini in the Real- ..., and Microsoft's coolant demonstration involved computational discovery followed by months of physical synthesis Transforming R&D with agentic AI: Introducing Microsoft .... Furthermore, any prize-level discovery—such as a new high-temperature superconductor or breakthrough battery chemistry—will inevitably require rigorous independent human replication, making a hybrid human-AI workflow the path of least resistance.

Closed-Loop Autonomous Labs (23%) Autonomous closed-loop laboratories represent the strongest challenger. Finding new materials is often a high-throughput, empirical endeavor, and well-funded, dedicated facilities are scaling rapidly. Periodic Labs is aiming for around-the-clock superconductor discovery with its 1-trillion-parameter Periodic Neon model News Betting on AI and Robots to Automate Superconductor ... Periodic Labs, and Google DeepMind is building a Gemini-integrated UK lab for high-throughput characterization Google DeepMind partners with U.K. for its first automated .... Recent work by Lila successfully screened 2,942 catalysts, yielding a highly stable acid-OER candidate https://arxiv.org/abs/2609.30133 How an AI-run lab cracked open green hydrogen's catalyst .... However, physical experimental throughput remains constrained, growing only from roughly 21 to 32 published samples a day since 2023. Scaling up physical synthesis and validation remains a particularly severe bottleneck that autonomous discovery workflows must overcome. These loops also frequently still require human intervention to move samples or interpret ambiguous secondary phases Agentic LLM Reasoning in a Self-Driving Laboratory for Air ..., keeping this option's probability below the human-execution route.

General Model Reasoning from Theory (20%) While materials science prizes overwhelmingly reward physically realized materials, the sheer pace of software-driven reasoning capabilities makes a purely theoretical breakthrough highly plausible. AI systems have recently saturated frontier math benchmarks, solved major conjectures (such as Sogge's local smoothing), and claimed existence proofs for complex theoretical physics problems like Yang-Mills theory and QCD Working with academic communities to build scientific ... AI makes discoveries in mathematics, computer science .... A general model could theoretically deduce a groundbreaking condensed-matter mechanism—such as explaining cuprate pairing or predicting an entirely new topological phase—long before physical laboratory automation matures. I assign this a 20% probability, balancing the unprecedented trajectory of theoretical reasoning against the traditional experimental bias of materials science awards.

Specialist Orchestration (16%) and Data Mining (8%) A general system directing purpose-built specialist models captures 16% of the probability. Machine-learned interatomic potentials and generative crystal models (like MatterGen A generative model for inorganic materials design) are improving rapidly, with thermal-conductivity prediction errors falling dramatically. An LLM agent orchestrating these tools is a credible architecture, as seen when Lila's integrated agent outperformed fixed-policy optimization https://arxiv.org/pdf/2609.30133. Mining existing datasets is the least likely path at 8%. While agents like Kosmos have successfully identified previously unreported processing relationships in existing data https://arxiv.org/html/2511.02824v2, achieving a Nobel-caliber breakthrough usually requires de novo synthesis or entirely new measurements rather than purely retrospective analysis.

Uncertainties and Categorization The most significant uncertainty in this assessment is the blurring of boundaries across these hybrid workflows. A future landmark discovery could easily involve an LLM reasoning from literature, querying a machine-learned interatomic potential, and passing a candidate structure to an automated synthesis loop that humans ultimately characterize and replicate. The resolution depends on where the "decisive scientific contribution" is attributed. Because of this structural overlap, probability is distributed broadly across the board rather than concentrated on a single dominant pathway.