LUCAN 10B Physics MoE is an 11B parameter multi-domain scientific foundation model built on a hybrid Mamba2 state-space and 12-layer Sparse Mixture-of-Experts (MoE) architecture with a PhysicsConstrainedOutputHead. It models continuous scientific latents across fluid transport, chemical reactors, phase-change thermal physics, and molecular systems.
LUCAN-10B-Physics-MoE is a multi-domain scientific foundation model developed by Shodh AI Private Limited. The model includes a native PhysicsConstrainedOutputHead (TCPI) enforcing non-negativity and integral mass/flux conservation directly inside the forward pass.
Multi-Domain Scaling Over LUCAN-3B: Achieves a 9.86% reduction in validation Relative L2 error (1.1933 -> 1.0756) and a 6.01% reduction in held-out test Relative L2 error (1.1682 -> 1.0980), improving across all 5 physics domain groups (Compressible Fluids -12.71%, Reactors & Erosion -6.18%, 3D Phase-Field Solidification -14.26%, Molecular Conformations -7.34%, and Aerodynamic Shock Benchmarks -10.01%) with zero dropped tokens.
Inference Latency & Speedup (1× NVIDIA H100 80GB): Delivers a 37.63 ms median single forward pass latency, 5.98 seconds end-to-end execution across a 1,000-scenario parameter sweep (5.98 ms/case), and a 39.83× end-to-end speedup (0.0801 s vs. 3.1914 s) over the numerical Stefan phase-change solver.
Other
Shodh Ai research Team
Token Classification Model
PyTorch
Restricted
Science, Technology and Research
05/11/26 06:54:45
4.02 GB
To preview this file, you need to be a registered user. Please complete the registration process to gain access and continue viewing the content.
Other
© 2026 - Copyright AIKosh. All rights reserved.