Using AI to create incredibly efficient solar cells: Our investment in Cosmos Innovation
By Dror Berman and Josh Rapperport
Technological progress is often described in terms of platform shifts. Personal computing. The internet. Mobile. As these new paradigms gain steam, the innovation that follows changes the course of our society. Our core thesis — the Super Evolution — is focused on how platform shifts in data, compute and engineering converge to produce rapid progress in domains where innovation has been driven by human experts. We continue to be fascinated by the way this intersection of technological and scientific methods creates combinatorial magic.
Two of the most important platform shifts of the coming decades will be artificial intelligence and the energy transition to carbon-free power. It goes without saying the current rate of improvement in AI is one of the most exciting moments in modern technology. Similarly, the talent and capital flowing into the energy transition is only accelerating and will need to continue if we’re to avoid dangerous climate change.
We know AI can transform how science gets done (as Eric Schmidt recently laid out) and we know we need profound innovation in the way we produce electrons if our climate goals are to become reality. This intersection of AI and renewable energy is where our latest investment, Cosmos Innovation, is building. We’re excited to have been a part of the team’s journey from the very beginning as they apply the latest in AI to create a better, more efficient physical world.
In concrete terms, Cosmos Innovation is radically accelerating and improving the process by which semiconductor development can produce optimal performance. For semiconductors, like all processes of experimentation, development has historically been human-driven. Researcher’s design-of-experiment (DoE) is guided by their understanding of underlying science, and they make changes through each iteration of an experiment to push toward some desired performance. We can think of this process as traversing a design universe, essentially a matrix of possible process conditions and material compositions, or a ‘recipe’, in the hopes of arriving at the optimal design. The various process parameters provide the ‘knobs’ that the researchers can turn, and they can only change one or two at a time.
Cosmos Innovation is using machine learning/iterative learning to turn all the knobs at once. In other words, they are using AI-driven experimental design to enable optimized outcomes with far fewer experiments, and much better target performance. They have demonstrated these capabilities with some of the largest semiconductor companies in the world, and the results are extraordinary — Cosmos Innovation can accelerate process optimization by 10x. The technology allows semiconductor manufacturers to develop novel, optimized recipes in a fraction of the time and cost of conventional methods.
Which brings us to the critical application of their technology — producing the most efficient solar panels ever made. Specifically, unlocking the full potential of perovskite-silicon tandem solar cells, a uniquely promising but not-yet-commercialized design. Perovskite-silicon tandem is the ideal candidate for AI-guided design because it is incredibly complex. While the underlying science and performance of crystalline silicon (over 90% of today’s global solar market) is well understood, adding a perovskite layer on top creates significant challenges. Perovskite recipe optimization is hard because of very high combinatorial complexity, somewhere around 5^72 possible permutations. Perovskite tandem architectures can have 12 or more layers, and over 100 input knobs, and there are no performant, physics-based models that researchers can rely on. Furthermore, the “goodness” factors, or relationships between the parameters driving ultimate performance, are also not well understood. This has led to fundamental challenges around degradation of produced solar cells, with stability and efficiency deterioration at elevated temperatures and humidity. In short, making sense of the overwhelming amount of combinations is impossible — it’s clear that human-driven approaches are too slow and won’t yield optimal results.
Why is it so important that perovskite-silicon reaches its full potential, and why is the market opportunity so big? Because of the incredible efficiency potential of perovskites. Delivering cheap, abundant power (often described as levelized-cost-of-energy or LCOE) from the sun requires optimizing two primary levers — solar cell cost and solar cell efficiency. While the last decade has seen an order of magnitude decline in costs, crystalline silicon efficiency has increased by only a few percent, going from roughly 18% to 25%, with a theoretical limit of 29%. Building on all the gains of silicon to date, the theoretical limit of perovskite-silicon is 43%. This remarkable potential is why the solar industry is so excited about perovskites.
The timing is ideal. Fundamental research in perovskites has been gaining momentum for the last decade, with record efficiencies nearing 34%. The fundamental science has been largely derisked, and now Cosmos Innovation is ready to use AI to close the gap to reality. Not only is the team addressing major challenges in designing solar cells, but they are also bringing self-learning to the manufacturing stage. As First Solar showed, it takes incredible execution and manufacturing prowess to produce novel solar architectures at scale. In this stage, AI can also be a game-changer. Cosmos Innovation intends to utilize next-gen capabilities to drive producibility, quality control and root-cause analysis, allowing for repeatable production of cells with commercial stability and performance. Automated, self-learning process control and tuning — long the ideal among visionaries in solar and semiconductors — has been largely out of reach over the last 50 years of solar manufacturing. The first version will be ready for experimentation beginning this fall at the Singapore-based fab.
We first partnered with Cosmos after hearing incredible things about the founders from Demis Hassabis, the CEO of DeepMind and Tomaso Poggio, the renowned MIT academic. It became clear to us that the founders, Vijay and Joel, are world-class technologists bringing together decades of AI, semiconductor, and solar expertise. Vijay led the AI effort at the Institute for Infocomm Research at the Agency of Science, Technology and Research (A*STAR), where he managed a portfolio of more than 50 AI projects across over 10 domains with a key focus on semiconductors. Joel was the former group head at the Solar Energy Research Institute of Singapore, where he led the development of an award-winning PV manufacturing technology. He also served as an AI team lead at A*STAR, where he led the development and deployment of AI solutions for Tier-1 manufacturers and R&D institutes in the semiconductor, material, and chemical domains. It’s rare to see a team of such multidisciplinary depth, and we’re proud to be supporting them alongside Xora and Two Sigma.
Solar energy is the single most impactful lever available to reduce planetary warming in this decade. That is the conclusion of a highly anticipated report from the Intergovernmental Panel on Climate Change (IPCC), which assessed the various mitigation pathways to curb climate change between now and 2030. AI provides an unprecedented opportunity to meet this moment and change the way we produce electricity. We look forward to the journey!
