Rohan Thorat
About | Portfolio | Blogs | Videos | GitHub | Publications | Teaching | Extra-activitiesResearch & Projects
Featured
Hover (or touch & hold) the ring to pause · full details below
Structural models always carry modelling assumptions — wrong beam theory, misspecified damping, incorrect boundary conditions. This framework uses a Gaussian Process Latent Force Model (GPLFM) in the reduced modal domain to learn and correct these discrepancies on-the-fly. A mesh-invariant neural network maps modal states to error estimates, validated across five realistic error scenarios including unmodelled nonlinearity and local damage.
Training RL directly on a physical structure is risky — random exploration can cause damage. This work proposes a hybrid LQR+RL framework where an LQR controller (derived from arbitrarily chosen, potentially wrong parameters) linearly mixes with the RL policy during training to keep exploration safe. After training, the RL policy takes over and surpasses LQR in vibration suppression with lower control effort — all without ever identifying the true system model.
Hyperelastic materials like rubber exhibit highly nonlinear stress-strain behaviour that is hard to model from first principles. This work uses a sparsity-promoting variational Bayesian framework to automatically select the right constitutive terms from a large candidate library — recovering the governing law with full uncertainty quantification on each coefficient, enabling downstream reliability assessment.
Starting from the REINFORCE policy gradient, this blog rigorously derives the full Actor-Critic algorithm step by step — reward-to-go, baseline subtraction (with no-bias and variance-reduction proofs), and the advantage function A(s,a). Each step is motivated by the problem it solves, so the reader builds genuine intuition rather than just following equations.