AlphaGeo is a multi‑agent pipeline for automating constitutive modelling: it configures models, performs parameter identification, and calibrates to data with minimal manual effort. The workflow: (1) Explorer samples parameters and initial conditions within bounds; (2) Self‑Play builds a surrogate dataset from states S and deltas D; (3) Learner trains a mapping f: X→ΔA to predict parameter updates; (4) MCTS refines ΔA to minimize calibration error.
Animation guide: The left panels show the Explorer’s bounds and Self‑Play trials; the center panel tracks Learner training; the right panel animates MCTS iterations. The final panel compares baseline, final prediction, and observations with a legend.
Explorer samples A (model parameters [a₁..a_M]) and G (initial conditions [g₁..g_NIC]) within their bounds. Each sample yields a prediction yᵢ and a delta Δyᵢ = yᵢ − y₀ (vs baseline y₀). States are S = {(x, y₀)} and deltas are D = {Δyᵢ}; together they form X = {S, D} for the learner.
X collects states S and differences D where S = {(x, y₀)} are baseline input‑output pairs, D = {Δyᵢ} are per‑trial prediction deltas; A are model parameters [a₁..a_M]; G are initial conditions [g₁..g_NIC].