Research

Decades of materials research have produced synthesis routes, characterization methods, and chemical intuition that no algorithm can replace. We build on this human expertise and add robotics and AI only where they genuinely accelerate the work.

Our robotic platforms are designed around the dexterity and experience of human researchers, so that automation reproduces proven experimental practice rather than reinventing it. We study AI-driven decision making—choosing the next experiment and learning from its result—and develop systematic ways to incorporate human domain knowledge into this learning loop, so that the model’s decisions are refined by expert insight rather than by data alone.

Our goal is a laboratory in which human, AI, and robot each do what they do best, and together accelerate the development of materials and devices.

A decarbonization technology is, at its core, a high-dimensional optimization problem. From material synthesis parameters and the resulting active materials, through the microenvironment around them, to system parameters and device operating conditions, dozens of coupled variables shape the final performance in practice. Exploring this space one factor at a time is slow, and the resulting discovery gap is a major reason why promising materials rarely reach deployment.

Our goal is to navigate this space far faster with the self-driving laboratory described above: robotic experimentation to generate data at scale, AI to decide the next experiment, and human domain expertise to shape the platform and refine those decisions. Along the way, the accumulated experimental data let us build a bridge between computation and experiment, which have largely developed in parallel.

In pursuing this, we build a self-driving lab that is both performance-driven and knowledge-aware. Our robust platform explores two design axes together. The first is materials-level design—defect engineering, porous structure, morphology and size, and crystalline structure—and the physico-chemical properties they give rise to. The second is device-level operating conditions—temperature, pressure, and the micro-environment around the active material—which decide the final performance in practice yet have rarely been treated as tunable variables in self-driving labs.

By generating end-to-end datasets that link synthesis, structure, properties, and device performance, the platform serves two goals at once. Performance-driven: optimize for high performance at scale, not only under fixed laboratory conditions. Knowledge-aware: uncover the underlying design principles that govern both axes, with the principles emerging in the loop rather than after it.