Best Outgoing Project • 2022–23 🏆

Physics-Informed Neural Networks (PINNs)

BMS College of Engineering

⚡Purely data-driven physics simulation required large labeled datasets expensive or impossible to generate experimentally; sparse training data produced physically implausible solutions - the model could satisfy the data loss while violating governing equations; no unified framework existed that validated across multiple physics domains simultaneously.

⚠️Physics simulation (fluid dynamics, structural mechanics, heat transfer) is unstable under purely data-driven approaches - requires large labeled datasets that are expensive or impossible to generate, and produces physically implausible solutions under sparse data.

⚙️Dual-loss PINN framework embedding governing PDE/ODE constraints directly into the optimization objective alongside data loss. Validated across six benchmarks: Burgers' equation, 1D heat conduction via pin fin, fixed-fixed column deflection, cantilever tip deflection, 1D transient cooling under Neumann flux and Dirichlet boundary conditions.

🛡️Physics constraints act as a regularizer - preventing physically implausible solutions from satisfying data loss alone. Neumann and Dirichlet boundary condition variants validated generalizability across constraint types and problem geometries.

🚀Stable convergence across 6 physics benchmarks - fluid, structural, thermal - with limited labeled data. Applied use cases in HVAC thermal feedback and server cooling. Best Outgoing Project - BMSCE 2022–23.

↳ Best Outgoing Project • BMSCE 2022–23 • 6 validated benchmarks across fluid, structural, and thermal domains

PINNsDual-Loss OptimizationPDEs / ODEsFluid DynamicsStructural MechanicsHeat Transfer