AI-Fitter: LLM-Assisted Curve Fitting for Physics Research Without Manual Initialization

A viable artificial intelligence (AI) physicist must perform three tasks with minimal human intervention: (1) form hypotheses, (2) design and run experiments, and (3) analyze data. In practice, curve-fitting for physics data analysis still hinges on hand-crafted choices (i.e. initial guesses, loss definitions, optimizer settings, and stopping rules) that are brittle across equations and noise regimes. To address this shortcoming, we introduce AI-Fitter, a large language model (LLM)-powered curve-fitting system that automates the entire fitting loop. Unlike toolkits tuned to narrow equation families, AI-Fitter configures the full loop for arbitrary analytic forms. On a benchmark of 120 physics equations spanning linear and nonlinear models with varied noise, it attains normalized root-mean-square error (NRMSE) < 0.1, often considered a good-fit threshold, in a substantial fraction of cases. Ablations further quantify the contribution of LLM reasoning to fit quality. As a proof of working principle, AI-Fitter is used to analyze resonance-frequency responses of a probe used in Shear-force Near-field Acoustics Microscopy (SANM) and extract the probe's viscoelastic properties needed to optimize SANM image acquisition. The results are verified with in situ applied standard methods obtained through a more time-consuming procedure. Overall, AI-Fitter reduces human effort in physics data analysis and advances the vision of a fully autonomous "AI physicist."