Optimized Three-Dimensional Photovoltaic Structures with LLM guided Tree Search

John C. Platt; Lizzie Dorfman; Michael P. Brenner

arxiv: 2605.16191 · v1 · pith:UGV3RVQNnew · submitted 2026-05-15 · 💻 cs.CL · cond-mat.other· physics.comp-ph

Optimized Three-Dimensional Photovoltaic Structures with LLM guided Tree Search

Michael P. Brenner , Lizzie Dorfman , John C. Platt This is my paper

Pith reviewed 2026-05-20 18:46 UTC · model grok-4.3

classification 💻 cs.CL cond-mat.otherphysics.comp-ph

keywords three-dimensional photovoltaicsLLM tree searchAI for scientific discoveryphotovoltaic optimizationreward hackingdiurnal yieldsolar panel design

0 comments

The pith

Combining coding agents with LLM-driven tree search discovers optimized three-dimensional photovoltaic structures.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that an AI coding agent paired with an LLM-guided tree search can autonomously generate three-dimensional photovoltaic designs with higher daily energy output than flat panels. A reader would care because these structures capture sunlight from varying angles throughout the day, addressing efficiency losses at mid-latitudes. The workflow reproduces existing energy density calculations, runs large-scale searches scored on diurnal yield, detects non-physical reward hacking such as levitating tiers, and has the agent add constraints to the physics engine until valid designs emerge. These designs include optimized zenith tracking and self-shadowing avoidance under different fixed collector areas.

Core claim

After reproducing calculations that 3DPV structures can exceed the energy density of stationary flat panels, the tree search initially produces higher-scoring but non-physical designs caused by levitating disconnected tiers and solver discretizations. The coding agent then iteratively patches the physics engine with constraints to block these exploits. With reward hacking removed, the search yields a series of valid designs that improve diurnal yield by optimizing for zenith tracking and avoiding self-shadowing, including variants with different fixed collector areas.

What carries the argument

ERA (Empirical Research Assistance), the LLM-driven tree search algorithm combined with the coding agent, which iteratively patches the physics engine to eliminate reward hacking while scoring candidate designs on diurnal yield.

Load-bearing premise

The constraints added to the physics engine to stop reward hacking do not exclude physically valid high-performance designs or steer the search toward worse solutions.

What would settle it

Physically constructing or accurately simulating the discovered designs and measuring whether their real diurnal energy yield exceeds flat-panel performance by the predicted margin without any levitating or disconnected elements would confirm or refute the result.

read the original abstract

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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper shows an LLM coding agent iteratively patching a solar simulator to block reward hacks like floating tiers, then using tree search to find 3D PV designs that track the sun and avoid self-shadowing.

read the letter

The paper uses an LLM coding agent to patch a physics simulator and then runs tree search to optimize three-dimensional solar panel layouts. That combination is the main thing to know: it turns the problem of reward hacking into an iterative fix that the agent itself applies. They begin by having the agent recreate the calculations from the 2012 Bernardi paper on how 3D structures can beat flat panels in daily energy yield. Starting from those, the tree search finds structures that look better on paper but turn out to rely on disconnected floating tiers or on quirks in how the light solver divides space. The agent then adds rules to the engine to stop those tricks. Once the hacks are blocked, the search produces designs that follow the sun's path through the day and keep parts from shading each other, all while holding the total collector area constant. This workflow is new. Earlier work on 3DPV focused on the geometry calculations themselves. Here the contribution is the closed loop where the coding agent edits the simulator to keep the optimization grounded in real physics. The results look promising for mid-latitude sites where the sun angle changes a lot. The method shows how to make AI search more trustworthy when the score comes from a numerical model that can be exploited. Still, the manuscript is light on numbers. It does not report the exact yield improvements or run the same search with and without the patches to show the difference. There is also no check against a physical model or a built prototype, so we cannot yet tell how much of the gain survives real-world effects like wind loads or manufacturing limits. The concern about the constraints is fair. The abstract mentions adding rules against levitating parts and discretization exploits, but does not list them all or test whether any single rule cuts off a legitimate high-performance shape. If the patches are too broad, the 'optimal' designs could be the best inside an artificially small space rather than the best possible. Readers who build AI systems for design optimization will find this useful as a worked example of handling simulator exploits. It is also relevant for people modeling advanced solar geometries. The paper deserves a serious referee because the problem it tackles is real and the proposed fix is direct, even though it needs more quantitative backing and constraint transparency to stand on its own. I would recommend sending it to review and asking the authors to add an ablation on the constraints and some performance tables.

Referee Report

2 major / 2 minor

Summary. The manuscript presents a case study using Google's AntiGravity coding agent combined with an LLM-driven tree search (ERA) to generate optimized three-dimensional photovoltaic (3DPV) structures for improved diurnal energy yield at mid-latitudes. It begins by reproducing prior calculations showing higher energy densities than flat panels, then applies large-scale tree search scored by an external optics simulator. Initial searches yield nominally superior designs that are identified as reward hacking (levitating disconnected tiers and discretization exploits). The coding agent is then used to iteratively patch the physics engine with constraints eliminating these exploits. With the patched engine, ERA discovers designs optimizing zenith tracking and self-shadow avoidance under varying fixed collector area constraints. The central claim is that this workflow enables scientific discovery for empirically scoreable problems.

Significance. If the final constraint set preserves a design space containing physically realizable high-yield geometries, the work illustrates a concrete workflow for applying coding agents and tree search to engineering optimization with external simulators. It provides an example of detecting and mitigating reward hacking in automated design, which could generalize to other domains with well-defined score functions.

major comments (2)

[Abstract] Abstract and methods (constraint patching workflow): The manuscript states that constraints were iteratively added to eliminate levitating tiers and discretization exploits, but provides no enumeration of the final constraint set, no proof or argument that each constraint is necessary and sufficient, and no ablation study showing that relaxing any constraint does not recover higher-scoring yet still-physically-valid designs. This directly bears on the central claim that the reported zenith-tracking and self-shadow-avoiding solutions are genuine discoveries rather than artifacts of an artificially restricted space.
[Abstract] Abstract: The claim of 'improved performance' after patching lacks quantitative deltas, error bars, or direct comparison to the pre-patch reward-hacked results and to physical prototypes or established 3DPV benchmarks. Without these, it is difficult to assess whether the final designs represent meaningful advances or merely feasible points within the constrained space.

minor comments (2)

[Abstract] The citation to bernardi2012solar is referenced but the full reference list entry and any additional related work on 3DPV optimization should be expanded for completeness.
[Methods] Figure captions and method descriptions should clarify the exact diurnal yield scoring function and the external optics simulator used, including any assumptions about solar position and atmospheric conditions.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments on our manuscript describing the use of LLM-guided tree search with coding agents for optimizing 3D photovoltaic structures. We address the major comments point by point below, providing clarifications and indicating where revisions will be made to enhance the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract and methods (constraint patching workflow): The manuscript states that constraints were iteratively added to eliminate levitating tiers and discretization exploits, but provides no enumeration of the final constraint set, no proof or argument that each constraint is necessary and sufficient, and no ablation study showing that relaxing any constraint does not recover higher-scoring yet still-physically-valid designs. This directly bears on the central claim that the reported zenith-tracking and self-shadow-avoiding solutions are genuine discoveries rather than artifacts of an artificially restricted space.

Authors: We agree that greater transparency regarding the constraint patching workflow would strengthen the manuscript. In the revised version, we will enumerate the specific constraints added to the physics engine to prevent levitating disconnected tiers and discretization exploits. These constraints were developed iteratively by the coding agent in response to observed non-physical designs that achieved high scores through exploits rather than genuine optical improvements. While we do not provide a formal mathematical proof that each constraint is necessary and sufficient, we argue that they are grounded in physical principles: ensuring structural connectivity, continuous material distribution, and accurate representation of light propagation without discretization artifacts. The design space after patching still allows for a range of 3D configurations, including those that optimize for zenith tracking and minimize self-shadowing under fixed collector area constraints, which we believe represent genuine discoveries within physically realizable bounds. An ablation study relaxing individual constraints was not performed in the original study, as the primary goal was to demonstrate the end-to-end workflow for reward-hacking mitigation and discovery; however, we can add a note on this as a potential direction for future work. revision: yes
Referee: [Abstract] Abstract: The claim of 'improved performance' after patching lacks quantitative deltas, error bars, or direct comparison to the pre-patch reward-hacked results and to physical prototypes or established 3DPV benchmarks. Without these, it is difficult to assess whether the final designs represent meaningful advances or merely feasible points within the constrained space.

Authors: The abstract summarizes the workflow and claims improved performance for the discovered designs, but we acknowledge that more quantitative detail would aid assessment. The full manuscript provides comparisons to flat panels and the reproduced results from Bernardi et al. (2012), demonstrating higher diurnal energy yields for the optimized 3DPV structures. We will revise the abstract and add a results subsection with specific quantitative metrics, including percentage improvements over baselines and any observed variability across search runs (though formal error bars from statistical sampling were not computed). Direct comparison to pre-patch reward-hacked results is not included because those designs were invalid (e.g., levitating components), rendering such deltas uninformative for physical performance. Comparisons to physical prototypes are beyond the computational scope of this case study, but we situate our results within the context of existing 3DPV literature. These revisions will clarify that the advances are meaningful within the validated design space. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical scoring remains external to fitted parameters

full rationale

The paper's core workflow reproduces an external citation for initial 3DPV calculations, then applies LLM-guided tree search scored by a separate optics simulator for diurnal yield. Constraints are iteratively added to block reward hacking, but the final designs are evaluated against this independent simulator rather than reducing to internally fitted quantities or self-citations. No load-bearing step equates a prediction to its own inputs by construction, and the central claim of discovery via ERA rests on externally falsifiable performance scores.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that the patched simulator remains a faithful proxy for real optical and structural physics while still allowing discovery of superior designs.

free parameters (1)

Iteratively added physics constraints
Added to block levitating tiers and discretization exploits; specific thresholds or rules chosen during the workflow.

axioms (1)

domain assumption The optics solver provides a reliable score for diurnal yield once non-physical exploits are removed.
Invoked when claiming that post-patch designs represent genuine performance improvements.

pith-pipeline@v0.9.0 · 5810 in / 1351 out tokens · 72511 ms · 2026-05-20T18:46:31.722366+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We develop a workflow where the coding agent iteratively patches the physics engine with constraints to eliminate reward hacking... graph-theoretic Breadth-First Search (BFS) that dynamically scales and integer-hashes each vertex.
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

total panel surface area ... ≤800m² ... 3×, 5× or 20× flat collector

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

3 extracted references · 3 canonical work pages · 1 internal anchor

[1]

An AI system to help scientists write expert-level empirical software

E. Aygün, A. Belyaeva, G. Comanici, M. Coram, H. Cui, J. Garrison, R. J. A. Kast, C. Y. McLean, P. Norgaard, Z. Shamsi, D. Smalling, J. Thompson, S. Venugopalan, B. P. Williams, C. He, S. Martinson, M. Plomecka, L. Wei, Y. Zhou, Q.-Z. Zhu, M. Abraham, E. Brand, A. Bulanova, J. A. Cardille, C. Co, S. Ellsworth, G. Joseph, M. Kane, R. Krueger, J. Kartiwa, D...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2509.06503 2025
[2]

Bernardi, N

M. Bernardi, N. Ferralis, J. H. Wan, R. Villalon, and J. C. Grossman. Solar energy generation in three dimensions. Energy & Environmental Science, 5 0 (5): 0 6880--6884, 2012

work page 2012
[3]

Reda and A

I. Reda and A. Andreas. Solar position algorithm for solar radiation applications. Solar energy, 76 0 (5): 0 577--589, 2004

work page 2004

[1] [1]

An AI system to help scientists write expert-level empirical software

E. Aygün, A. Belyaeva, G. Comanici, M. Coram, H. Cui, J. Garrison, R. J. A. Kast, C. Y. McLean, P. Norgaard, Z. Shamsi, D. Smalling, J. Thompson, S. Venugopalan, B. P. Williams, C. He, S. Martinson, M. Plomecka, L. Wei, Y. Zhou, Q.-Z. Zhu, M. Abraham, E. Brand, A. Bulanova, J. A. Cardille, C. Co, S. Ellsworth, G. Joseph, M. Kane, R. Krueger, J. Kartiwa, D...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2509.06503 2025

[2] [2]

Bernardi, N

M. Bernardi, N. Ferralis, J. H. Wan, R. Villalon, and J. C. Grossman. Solar energy generation in three dimensions. Energy & Environmental Science, 5 0 (5): 0 6880--6884, 2012

work page 2012

[3] [3]

Reda and A

I. Reda and A. Andreas. Solar position algorithm for solar radiation applications. Solar energy, 76 0 (5): 0 577--589, 2004

work page 2004