No full-body solve per gauge change: SHIFT-NVH holds median field error to 1.1% across 192 test cases
Synopsis
Luminary released SHIFT-NVH, a large physics model trained on finite-element solves of a single body-in-white that takes the gauges of 15 parts, the material and the vehicle speed and returns RMS acceleration in x, y and z at all 130,713 surface nodes under ISO 8608 class B road excitation. On 192 test cases split by design, with material and speed unseen, the median relative L2 field error is 1.1% and the median part RMS error is 0.47%, about one tenth of the 5.4% median gauge-induced change. The model covers a fixed mesh, ±30% gauge, three materials and four speeds from 30 to 110 km/h, and final gauge decisions still rest on full FE and physical testing.
Interpretation
Once body geometry is frozen, sheet gauge and material are the main variables left for noise, vibration and harshness (NVH), and each gauge change otherwise needs a modal and forced-response solve on the full-body finite element model. SHIFT-NVH replaces that step with one network forward pass, so engineers can evaluate gauge and material changes and see which parts each change affects without holding the FE model. Previously gauge, material and vehicle speed multiplied the candidates faster than an NVH team could solve them before gauges were released, and suppliers often lacked the full-body model or had to rework it for their toolchain. On 192 held-out cases the median relative L2 field error is 1.1% and the worst 2.6%, and the median part RMS error is 0.47% with a p95 of 2.4%.
The model maps design parameters to a response field on a known surface mesh, so it learns how thickness and material redistribute vibration rather than how geometry varies. It uses Luminary's Luminary-SMART encoder-decoder with 19 inputs, the 15 gauges plus Young's modulus, density, Poisson's ratio and vehicle speed, and outputs x, y and z RMS acceleration at each node, with part-level vibration taken as the node average over each part. The 1,212 cases split by design into 840 training, 180 validation and 192 test cases, with all four speeds and three materials of a design in the same split; training ran for 300 epochs on 4 NVIDIA H100 GPUs and the best validation relative L2 error was 0.020.
A second SMART model trained on the steel modal data takes only the 15 gauges and predicts the first 5 elastic mode shapes. That modal model does not predict frequencies, so it supplies shapes separately from the frequency solve. On 16 unseen steel designs the mean MAC against the reference is 0.98 to 1.00 for modes 1 to 4 and 0.92 for mode 5.
Perspective
The demonstrated setting is the 2010 Toyota Yaris body-in-white on a fixed mesh, with 15 parts varying by up to ±30% of nominal gauge, steel, aluminium and magnesium, speeds of 30, 50, 80 and 110 km/h and an ISO 8608 class B road; within that range engineers can screen gauge and material combinations, and full FE and physical testing remain the basis for final gauge decisions on shortlisted designs.
The model does not learn geometry variation and only learns how thickness and material redistribute the response over a known mesh; its accuracy figures come from 192 test cases spanning 16 unseen designs, 3 materials and 4 speeds, and behavior outside that gauge range, part set or road class is not given within the stated scope.