AI-driven closed-loop synthesis accelerates development of balanced cathode materials for batteries

A team from McGill University, Mila–Quebec AI Institute and Université de Montréal has demonstrated that a closed-loop combination of machine learning and robotic synthesis can navigate an enormous compositional landscape and deliver cathode materials with markedly improved, balanced performance. According to the authors’ report in Advanced Materials, their system evaluated roughly 14.2 million triple-doped variants of lithium cobalt phosphate (LiCoPO4) and, after fewer than 200 experimental syntheses, identified compositions that delivered a fivefold composite improvement across four key metrics: practical capacity, irreversible capacity, cycle retention and overpotential.

The workflow fused a pretrained set-transformer neural network with a multi-task Gaussian process to produce a surrogate model that predicts multiple electrochemical outputs together with uncertainty estimates. That model scored the entire 14 million candidate space in under 20 minutes on a single GPU; an acquisition function then selected small batches for synthesis by liquid-handling robotics and conventional sol–gel and sinter processing. Running three rounds of active learning, the team moved from an initial hit rate of about 8.8% for candidates beating the undoped baseline on all four metrics to 44.4% in the final round, signalling rapid convergence on reliably good chemistries.

Crucially for industrial deployment, the optimisation did not rely on explicit structural inputs; the model was trained on composition and performance data yet produced largely single-phase materials as verified by X‑ray diffraction, indicating dopant incorporation rather than formation of electrochemically inert second phases. The algorithm also diversified the palette of useful dopants, shifting away from an earlier reliance on indium toward chromium, niobium, scandium and phosphorus, potentially reducing supply‑chain concentration risks as scale-up is considered.

The study highlights the value of multi-objective optimisation in battery development. Previous efforts frequently targeted single properties and encountered trade‑offs that undermined holistic performance. By predicting and prioritising composite outcomes, the closed-loop approach produced candidates such as an Al–Cs–In combination that delivered near‑theoretical capacity with dramatically improved retention and lower overpotential, while also surfacing high performers that contained no indium at all.

This work sits alongside a growing body of AI-driven materials discovery that is beginning to reshape how developers in the energy sector search chemical space. According to an arXiv preprint describing the Energy‑GNoME database, AI pipelines can already identify tens of thousands of candidate materials for energy applications by mitigating data bias and constructing efficient feature spaces. Other machine‑learning frameworks, such as DRXNet, have been trained on thousands of experimental discharge profiles to predict full voltage curves across chemistries, aiding selection and optimisation of cathode families. Complementary studies reported in both the literature and industry blogs show rapid AI‑led identification of low‑lithium and magnesium cathode candidates, and initiatives by organisations such as SES AI, backed by NVIDIA tooling, aim to accelerate electrolyte and molecular discovery across vast chemical search spaces.

For industrial decarbonisation professionals the implications are practical. The McGill–Mila–Montréal demonstration reduces the experimental burden required to find multi‑objective improvements from thousands of trial syntheses to a few hundred targeted experiments, trimming time and resource costs in early development. The approach’s ability to prioritise supply‑chain‑resilient elements and to produce candidates that are structurally compatible with existing processing steps strengthens its appeal for scale-up evaluation and pilot production.

Caveats remain. The team reported at least one false positive whose apparent ultra‑low overpotential was revealed, upon manual review, to be an artefact of the electrochemical measurement. That underlines the need for domain expertise and quality control alongside automation. Moreover, successful laboratory optimisation must still clear subsequent hurdles, electrode engineering, manufacturing compatibility, lifecycle testing and cost modelling, before yielding commercial cathodes suitable for large‑scale electrification and heavy‑duty decarbonisation applications.

The authors position their surrogate model as broadly generalisable: it accepts arbitrary target properties and is not anchored to LiCoPO4 alone. Coupled with parallel developments in predictive voltage modelling, discovery databases and accelerated DFT workflows, the result is a maturing toolkit that industry R&D groups can integrate to shrink lead times for new battery chemistries, potentially accelerating the replacement of fossil‑fuel‑dependent equipment with electrified alternatives that support decarbonisation targets.