ML Research · NIT Calicut · 2025

PINN · Seismic Response

Physics-informed neural network predicting the seismic response of steel braced frames.

Physics-informed neural network in TensorFlow predicting the seismic response of steel braced frames.

My role: ML intern
Links: GitHub (planned) ↗
Stack: PythonTensorFlowNumPy; Python · TensorFlow · NumPy

92%

Prediction accuracy on test data

Source · Illustrative simulation of a single-storey response (Newmark-β), not the internship dataset

01 — Problem

Finite-element simulation is accurate but slow, while purely data-driven models can violate physical laws.

02 — Approach

01
Data loss + physics residual from the equation of motion.
02
Evaluated on unseen ground motions.

03 — Architecture

System architecture

S01
Ground motion
S02
Network
S03
Physics residual
S04
Loss
S05
Response

04 — Engineering notes

Physics inside the loss

The network is penalised both for missing the data and for violating the equation of motion m·ü + c·u̇ + k·u = −m·a_g at collocation points.

python

# Simplified sketch of the training loss (TensorFlow)
with tf.GradientTape(persistent=True) as g2:
    g2.watch(t)
    with tf.GradientTape() as g1:
        g1.watch(t)
        u = model(t)
    u_t = g1.gradient(u, t)
u_tt = g2.gradient(u_t, t)

residual = m * u_tt + c * u_t + k * u + m * a_g(t)
loss = mse(u_obs, model(t_obs)) + lam * tf.reduce_mean(tf.square(residual))

Result: 92% prediction accuracy on held-out ground motions.

Outcome

92% prediction accuracy with physically consistent outputs.

← All projects

Next project

Safira AI Chatbot