Engine Brain
Latent Physics Model
A learned model trained on high-fidelity solver outputs. It maps input parameters directly to physics fields — temperature, stress, velocity, pressure — in a single forward pass, without iterating a traditional solver.
How the LPM Works
Traditional solvers iterate mesh equations until convergence. The LPM learns the mapping from parameters to solutions, replacing iteration with inference.
Solver Data Generation
High-fidelity simulations across parameter sweeps produce labeled (input → field) training pairs using FEM, LBM, FDTD, and Phase Field solvers.
Latent Encoding
A physics-informed encoder compresses full field solutions (temperature, stress, velocity, pressure) into a compact latent space that preserves conservation laws.
Optimizer & Registry
Hyperparameter optimization with model versioning. Every trained LPM is stored with its solver lineage, training domain, and validated accuracy bounds.
Instant Inference
At query time, the LPM maps input parameters → latent → decoded physics fields in a single forward pass. No iterative solving, no mesh convergence loops.
Capabilities
Trained Domains
Thermal, Structural, CFD, EM
Each domain has its own trained checkpoint with solver-verified accuracy
Inference Speed
~1ms per query
Single forward pass on GPU vs minutes-to-hours for traditional solvers
Physics Consistency
Hard constraints
Conservation laws enforced in the architecture, not just the loss function
Hallucination Detection
Built-in
Transformer module flags outputs that violate physical bounds
Source Modules
genesis/lpm/architecture_pro.pyNeural architecture: encoder, latent space, decoder with skip connections and physics-aware normalization layers.
genesis/lpm/optimizer.pyTraining loop with learning rate scheduling, early stopping on physics residual, and multi-objective loss (data fidelity + PDE residual + boundary).
genesis/lpm/registry.pyModel versioning and retrieval. Stores trained checkpoints with metadata: solver source, training domain, accuracy bounds, parameter ranges.
How LPM fits into Genesis
Solvers generate data
FEM, LBM, FDTD, Phase Field
LPM learns the physics
Latent encoding + hard constraints
Instant inference
~1ms queries for real-time design
Explore the full platform
The LPM is one layer of the Genesis Engine. See the full solver library, multiphysics couplers, and GPU backends.