Optimized Three-Dimensional Photovoltaic Structures with LLM guided Tree Search
Michael P. Brenner, Lizzie Dorfman, John C. Platt
An LLM-driven tree search system called ERA (Empirical Research Assistance), combined with Google's AntiGravity coding agent, autonomously designs 3D photovoltaic structures that outperform flat solar panels at mid-latitudes by finding geometries that maintain favorable sun angles throughout the day. The key technical contribution is the reward-hacking mitigation loop: when the search found physically impossible designs (levitating tiers, discretization exploits) that scored artificially high, the coding agent iteratively patched the physics engine with constraints to close those loopholes, after which ERA found genuinely improved designs across several constraint regimes.
Google Research has published the ERA codebase directly at google-research/era with this solar optimization as a featured case study, so the infrastructure is real and accessible. Zero citations so far and the paper is brand new โ it's more a demonstration of the ERA platform's scientific discovery workflow than a solar engineering advance per se, meaning the relevant signal for builders is the agentic loop architecture (score function โ tree search โ reward hacking detection โ physics engine patching), not the photovoltaic results themselves.
We present a case study for how AI coding systems can be used to generate novel scientific hypotheses. We combine a generic coding agent (Google's AntiGravity) with an LLM-driven tree search algorithm (Empirical Research Assistance / ERA) to autonomously generate high-efficiency three-dimensional photovoltaic (3DPV) structures that overcome losses limiting flat solar panels at mid-latitudes. These structures operate by presenting favorable angles to the sun throughout the day, and for illustrative purposes we focus on optimizing performance for a single solar day. Our workflow begins by using AntiGravity to reproduce calculations \cite{bernardi2012solar} showing that 3DPV can have energy densities much higher than stationary flat PV panels. We use these initial designs as the starting point for large scale tree search, where we seek improved solutions and score them for their diurnal yield. The initial tree search leads to nominally more efficient solutions, yet they are caused by algorithmic reward hacking, arising from non-physical design features such as structurally levitating disconnected tiers and exploitations of the discretizations in the optics solver. To counteract this, we develop a workflow where the coding agent iteratively patches the physics engine with constraints to eliminate reward hacking. With reward-hacking eliminated, ERA discovers a series of designs with various constraints and improved performance, including optimal designs with different fixed collector areas, optimizing zenith tracking and avoiding self shadowing. Combining coding agents with tree search (ERA) provides a powerful platform for scientific discovery, for problems whose solutions can be empirically evaluated with a score function.