Machine Learning with Science

Physics-informed models, neural networks for scientific prediction, and computational tools for real-world physical systems.

I pair physics intuition with rigorous experimentation and modern deep learning to model real phenomena. From edge-ready vision architectures to physics-informed neural networks and scientific regression systems, I design pipelines that stay faithful to the underlying science while meeting production constraints.

Projects

Scientific machine learning work spanning physics-informed models, computational tools, and deployable ML systems.

Neural Architecture Search & Physics-Informed Solar Models

Thesis + Ibero Research | 2023 - 2024

Edge vision architectures discovered via NAS and a physics-informed neural network for thin-film solar cells - one track focused on latency and accuracy for deployment, the other on domain-grounded performance gains.

  • NAS workflow surfacing compact CNNs tuned for latency/accuracy and deployable on Raspberry Pi with ONNX Runtime.
  • End-to-end pipeline for edge vision: dataset creation, augmentation, compression, pruning, and quantization with benchmarks across accelerators.
  • 4-layer PINN (100-32-16-8) reaching MAE = 0.00065 and R2 = 0.92 on solar absorption predictions.
  • Raised absorption AUC by 112% (0.09 -> 0.73) with Flask API and interactive UI for real-time inference.
  • Automated 1,710 curve comparisons with MSE-based Low/Medium/High thresholds; earned a top 5% research scholarship.
NAS Edge AI Vision Models ONNX Runtime Model Compression PINNs Renewable Energy TensorFlow Scientific ML Optimization
View solar PINN paper

CO2 Mass Transfer Coefficient Prediction

Scientific ML research | Peer review in progress

Scientific ML model predicting CO2 mass transfer coefficients in packed absorption columns with high fidelity.

  • R2 = 0.99 and MAE = 2.437e-05 with a deep neural network.
  • Dual-activation design: ReLU early, Sigmoid deeper for better mapping.
  • Adaptive optimizer tuned architecture, batch size, and LR to sub-5e-5 MAE across 138 process conditions.
  • Early stopping and validation strategies to maintain generalization.
  • Findings documented in a paper under peer review.
Scientific Machine Learning Chemical Engineering Neural Networks Regression Models
View paper

Total Energy Neural Network

GitHub repository

Neural regression model predicting total energy (Ry) from mixed numerical and categorical features with reusable preprocessing.

  • Preprocessing pipeline with one-hot encoding and scaling feeding a Keras model.
  • Architecture: input aligned to processed features, two hidden ReLU layers, linear output.
  • Reusable assets: preprocessor.joblib and my_model.h5 for deployment.
  • R2 = 1.00, MSE = 0.81, MAE = 0.61, RMSE = 0.90.
  • Instructions for loading artifacts and predicting on new data.
Neural Networks Regression Keras Scientific Data Feature Engineering
View on GitHub

Einstein Field Equations Computational Tool

Mathematica program

Mathematica toolkit that computes full differential-geometry tensors and Einstein field equations for arbitrary coordinate systems and metric tensors.

  • Computes Christoffel symbols, Riemann tensor, Ricci tensor, Ricci scalar, Einstein tensor.
  • Accepts arbitrary coordinate systems with user-defined metric tensors.
  • Automates tensor algebra workflows used in general relativity.
  • Includes energy-momentum tensor support for perfect fluids.
  • Built for physicists, researchers, and students studying GR.
General Relativity Mathematica Tensor Calculus Differential Geometry