HaomingLuo/AgentFEM-Controlled-Transient-Heat

AgentFEM T3 Controlled Transient Heat

This dataset contains 192 complete finite-element trajectories from 64 independent physical scenarios. A two-dimensional metallic solid is heated by three localized volumetric sources and cooled by convection along its exterior. Every scenario branches from the same state at 300 s into three possible futures: maintain source power, throttle all sources to 40%, or shut down.

Why the branches matter

An ordinary transient dataset shows what happened under one command sequence. T3 also stores what would have happened from the same physical state under alternative commands. It therefore supports full-field forecasting, hotspot prediction, reduced-order modeling, operator learning, state estimation and model-predictive-control research.

Data

Start here: compare one scenario's three branches after 300 s. Can a predictor distinguish continued heating, reduced power and shutdown from the same initial field? Keep each complete triplet together when splitting the data.

AgentFEM heat-transfer guide · Introductory heat exercises

• 64 independent combinations of conductivity, density, heat capacity, convection and three source layouts;
• 3 counterfactual branches per scenario;
• 61 accepted states over 0–600 s;
• one fixed 36×24 quadrilateral mesh with 925 CG1 temperature degrees of freedom;
• full temperature fields, three controls, five sensors, hotspot coordinates, heat content, applied heat rate, outward heat rate and balance residual;
• SI units throughout.

cohort.h5 stores the complete arrays, index.csv and index.jsonl provide trajectory-level metadata, quality.json records the scientific gates, and preview.png shows one complete counterfactual triplet.

Array	Shape per trajectory	Unit / meaning
temperature_history_k	61 × 925	nodal temperature, K
source_control_w_m3	61 × 3	Gaussian-source peak power density, W/m³
sensor_temperature_k	61 × 5	fixed virtual sensors, K
hotspot_coordinates_m	61 × 2	hottest nodal location, m
thermal_content	61	sensible thermal content per unit depth, J/m
applied_heat_rate	61	applied heat rate per unit depth, W/m
outward_heat_rate	61	outward heat rate per unit depth, W/m
heat_balance_residual	61	discrete balance residual per unit depth, J/m

The sampled ranges are conductivity 120–220 W/(m·K), density 2500–3200 kg/m³, specific heat 700–1000 J/(kg·K), convection 10–40 W/(m²·K), source width 4–9 mm and source peak power density 2–8 MW/m³. These are bounded synthetic descriptors, and every realized value and source location is stored in the trajectory metadata.

Run a three-branch comparison

Download code/compare_branches.py, then run:

python -m pip install numpy matplotlib h5py huggingface_hub
python compare_branches.py

The script downloads the pinned 77 MB cohort, verifies all 64 triplets have identical temperature fields and controls through 300 s and remain within one split, then plots the first test scenario. No FEM runtime or GPU is needed. If you already have the data, use --data /path/to/cohort.h5; choose another group with --scenario SCENARIO_ID.

Read the three panels together: source commands change after 300 s, while temperature and stored heat respond over time. The command panel uses the recorded source controls, not the applied_heat_rate ledger, which is not synonymous with source-only power. Stored heat is plotted relative to the initial value and per unit out-of-plane depth. These are synthetic FEM trajectories, not model predictions or experimental control trials.

Leakage-safe split

The 64 scenario groups are split 48/8/8 into training, validation and test. Their 192 trajectories are therefore 144/24/24. All three branches from one physical scenario remain in the same split, so an identical pre-intervention field never appears on both sides of the evaluation boundary.

Verification

• 192/192 trajectories completed with finite fields;
• largest cohort heat-balance relative residual: 1.409e-11;
• maximum temperature/control difference before branching: exactly zero;
• manufactured-solution refined relative L2 error: 0.2605%;
• time-step halving changes final sensor values by 0.2415%;
• mesh doubling changes final sensor values by 0.0023%.

Baseline

A validation-selected PCA/ridge autoregressive state-space model obtains test full-field RMSE 9.747 K, normalized temperature-rise RMSE 19.72%, and maximum-temperature MAE 6.207 K. Persistence gives 49.420 K. The baseline is deliberately simple and defines a reproducible lower bar rather than a claimed best model.

Scope

T3 v1 is a controlled solid-conduction benchmark. It uses constant isotropic thermal properties within each trajectory, a synthetic metallic parameter range and unit out-of-plane thickness. It does not include radiation, phase change, fluid flow or thermo-mechanical feedback. Those effects belong to later, separately versioned datasets.