A new preprint review highlights physics-informed Bayesian optimization (BO) as a way to bring scientific knowledge into self-driving materials laboratories. In plain terms, BO is a closed-loop approach: the laboratory uses what it has learned to help select the next experiment. The review describes domain knowledge entering through representations, prior assumptions, mathematical kernels, acquisition functions—the rules used to rank next steps—and constraints.
One cited example gives the review’s clearest numerical contrast. Physics-informed BO approached a theoretical maximum transformation temperature of approximately 340 K in about 15 iterations, while black-box BO did not reach that value within 30 iterations. The review does not report a pooled uncertainty estimate for this comparison.
A search across many kinds of materials
This is a review of methods and representative applications, not a new experimental campaign. It covers the BO loop, surrogate and acquisition design, practical complications and physics-informed approaches. The authors’ focus is on how these pieces can be used in closed-loop workflows under realistic experimental challenges.
The examples span semiconductors, catalysts, chemical reactions, batteries, alloys and quantum materials. Across those areas, the review describes representative advances associated with thin-film synthesis, catalysis, photocatalytic reactions, lithium-metal batteries, Invar alloys and Weyl-fermion materials.
From reaction mixtures to battery cells
One chemistry example puts the scale of the search in concrete terms. In a photoredox campaign, 55 catalysts were tested from a virtual library of 560 molecules, while 107 of 4,500 possible formulations were evaluated. The maximum cross-coupling yield increased from 71% to 88%; random sampling reached 75%.
A lithium-metal battery example reported a roughly threefold increase in mean symmetric-cell lifetime after three learning iterations. The five highest-ranked formulations had mean cycle lives of approximately 200 to 235 cycles, compared with approximately 60 to 185 cycles for three literature-reported benchmark formulations.
A separate full-cell task reported average capacity retention after 100 cycles of 58.2% for the initial formulations and 84.0% for the second-round top formulations. The review presents these as outcomes from separate applications, not measurements from one shared battery experiment.
The targets range from thin films to quantum transport
Other examples show how different the targets can be. In a semiconductor case, human-executed re-optimization further reduced the Urbach energy to 163 meV. In catalytic materials, active learning began with 30 experimentally characterized oxides, selected 10 additional compositions, and identified CPCF with an intrinsic overpotential of 391 mV, the lowest among the 40 perovskite oxides evaluated.
In an alloy search, six iterations involved only 17 new alloys from a space containing millions of possible compositions. The campaign identified two high-entropy Invar alloys with thermal-expansion coefficients of approximately 2 × 10^-6 K^-1 at 300 K.
Quantum materials supplied another demanding test. In a BO-optimized SrRuO3 workflow, a low-temperature regime described as quantum transport of Weyl fermions emerged above an RRR of approximately 20 and extended to about 10 K in the highest-quality films. The optimized films exhibited five transport signatures associated with Weyl fermions and quantum mobilities of approximately 104 cm2 V-1 s-1.
The difficult questions come later
The review’s larger argument is that BO should move beyond finding a promising setting and toward experiments that are reliable, interpretable and able to generate reusable scientific knowledge. It presents physics-informed components as a way to add expert knowledge while retaining data-driven updating, provided their assumptions are tested against simpler baselines.
That ambition leaves several questions open. The review identifies nonstationarity, multimodal observations, adaptive problem formulation and scientific reasoning by humans, large language models and research agents as unresolved problems.
The review brings together examples from different materials and fields rather than reporting one common experiment. Its numerical results are therefore illustrations from separate cited studies, not a single pooled estimate or proof that BO is universally better than simpler methods.
No new data were generated or analyzed as part of the review, and the authors declare no conflicts of interest.
Paper data and sources
Original title: Bayesian Optimization for Self-Driving Materials Laboratories: From Algorithms to Physics-Informed Workflows
Authors: Yuki K. Wakabayashi, Takuma Otsuka
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text