Luminary attaches calibrated 90% error bars to trained large physics models with a 166,000-parameter companion network

Luminary Cloud's Aashwin Mishra and Dheeraj Vemula describe an uncertainty method for models built on the Luminary-SMART architecture. The trained model is frozen; a companion network of about 166,000 parameters, roughly 1% of the model, takes its per-point features, a geometry summary and the prediction, and returns an error-bar size for each field. It is trained with an interval-score loss on the model's errors, then calibrated on held-out simulations with normalized split conformal prediction; drag and lift are calibrated separately. The post uses 90% coverage and reports results on SHIFT-Wing, SHIFT-SUV, SHIFT-Truck and DrivAerML: on SHIFT-Wing the drag bar is ±0.0014 (14 counts) and contains the CFD drag for 69 of 73 test designs, and on SHIFT-SUV the bars contain the CFD value at 90.0% of surface points on 74 held-out designs. Luminary states that the calibration does not extend to designs far from the calibration set, which its out-of-distribution score is meant to flag first.

Share on LinkedIn, X or Bluesky, or copy the link.