Bayesian experimental design for self-driving labs

Methods for choosing the next experiment, deployed in autonomous laboratories for materials science.

Finding the right experimental conditions for new materials is often like finding a needle in a haystack: vast design spaces, expensive experiments, and complex interactions between variables. We address this by developing Bayesian experimental design methods and deploying them in self-driving laboratories—autonomous systems that combine robotics and AI into a closed-loop for rapid experimentation and exploration.

Batteries

Rechargeable batteries are key for transitioning to a carbon-neutral future, but it will require huge quantities of battery materials. It is therefore of utmost importance that both the battery materials and the manufacturing process are sustainable and cost effective. In these regards, batteries based on aqueous electrolytes are especially promising. However, water has a very narrow electrochemical stability window, which poses a major challenge for aqueous electrolytes. In principle, it can be accommodated by including functional additives to the electrolyte, but finding the right “recipe” is akin to finding a needle in a haystack.

ODACell robotic setup

ODACell: a robotic setup for automated electrolyte formulation and coin cell assembly.

Together with the research group of Erik J. Berg at the Ångström Advanced Battery Centre we’ve developed ODACell, a robotic setup for automated electrolyte formulation and battery assembly, capable of preparing large batches of coin cells. The coin cells are automatically cycled to test their performance and new experiments suggested using Bayesian optimization for data-driven decision-making. In ongoing work, we are adding capabilities such as in operando sensing using Electrochemical Quartz Crystal Microbalance (EQCM).

Thin film science

In thin film science, self-driving labs have mainly been restricted to solution-based synthetic methods which are easier to automate but cannot access the broad chemical space of inorganic materials. In collaboration with the research group of Jonathan Scragg, in the division of solar cell technology, we’re building a self-driving lab based on magnetron co-sputtering. We present a rapid and calibration-free in-situ, machine learning driven approach to produce composition maps for arbitrary source combinations and sputtering conditions. Our framework dramatically increases throughput by avoiding the need for extensive characterisation or calibration, thus demonstrating the potential of self-driving labs to accelerate materials exploration.

Self-driving sputtering lab workflow

Simulations

In parallel with the physical experiments above, we are also performing simulation studies to better understand the underlying mechanisms. For instance, we recently studied electrode/electrolyte interface side reactions and dendrite growth in fast-charging metal-anode batteries, by combining a phase-field model of electrodeposition with Bayesian optimization. Our approach allows exploration of the parameter space with high sample efficiency and low computation complexity.

Simulated dendrite

Phase-field models of this kind have many physical parameters. Using Lie symmetry groups, we reduce the parameter space of such a model without changing its input-output behavior. The reduction gives exact mappings between parameter sets, for instance sets that produce identical dendrite morphologies in battery cells of different sizes.

Bayesian experimental design

Gaussian processes are the standard surrogate model for choosing the next experiment, and we study how they shape those choices. On a bounded design space, the posterior variance of a Gaussian process with a stationary kernel is inflated near the boundary, because the kernel’s correlation neighborhood is cut off there. This effect worsens with dimension and does not depend on the measured values. Its consequences differ between acquisition functions: maximizing the variance concentrates measurements in the corners, while integrated posterior variance and expected predictive information gain place them on axis-aligned shells in the interior.

For active learning, where the goal is to reconstruct an unknown function, we correct the boundary bias with a geometric equalizer. A second limitation is that the posterior variance depends only on where measurements are made, not on what they show. A warp of the input space, built from the posterior mean without any training, therefore puts more measurements where the function varies rapidly. For Bayesian optimization, we keep the surrogate fixed and set the spatial preference of the exploration term explicitly, through spatial weight fields whose construction is amortized by neural surrogates. The default boundary bias changes which optima are reached, and alternative exploration principles often lower the regret, particularly early in the optimization.