CAED-Agent: an Agentic Framework to Automate Simulation-Based Experimental Design

Yang Zhang, Yadi Cao, Sophia Huiwen Sun, Rose Yu

Keywords: Cost-Efficienct Experimental Design; Scientific Agents; In-Context Optimization

TL;DR: We present CAED-Agent, a large-language-model agent framework that utilizes inference-time scaling with lightweight surrogate feedback for cost-aware experimental design.

Abstract: Configuring physics-based simulations requires balancing granularity against computational budget, a dilemma we term **C**ost-**A**ware **S**imulation-Based **C**onfiguration **O**ptimization (CASCO). Traditional methods, such as Bayesian optimization or manual expert design, often struggle with the curse of high dimensionality or fail to generalize. Large Language Models (LLMs) offer promise for automating such workflows but, as we show experimentally, lack inherent cost awareness and frequently propose inefficient configurations. While inference-time scaling can improve the exploration width to find cost-efficient configurations, it demands prohibitively many simulator queries. We propose **C**ost-**A**ware **S**imulation **C**onfiguration **O**ptimization **Agent** (CASCO-Agent), an agentic framework guiding inference-time scaling via lightweight surrogates that predict low-dimensional metrics (accuracy, cost) rather than complete physics fields. This enables easier training and flexible adaptation to data availability, e.g., using Gaussian Processes in data-scarce regimes or Neural Networks when data is abundant. In experiments across 3 typical PDE solvers (elliptic, parabolic, and hyperbolic), CASCO-Agent consistently outperforms Bayesian optimization and LLM-based baselines, achieving success rates comparable to inference-time scaling with a ground truth simulator without incurring expensive simulation overhead.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 15143