Computational chemist and ML researcher at ICT Mumbai. I develop neural architectures and open-source tooling for vibrational spectroscopy — connecting group theory, information theory, and deep learning to solve the spectral inverse problem.
| Project | Description | Links |
|---|---|---|
| Spektron | Foundation model for vibrational spectroscopy — D-LinOSS backbone, masked pretraining on 222K spectra, VIB disentanglement, Sinkhorn OT calibration transfer. Targets 10-sample transfer where classical methods need 50. | Code · Page |
| ReactorTwin | Physics-constrained neural ODE digital twin for chemical reactors. Hard post-solve constraint enforcement (mass balance, positivity, stoichiometry) — exact satisfaction by construction, not soft penalty. 1500x faster than first-principles simulation. v1.1.1, 1509 tests. | Code · Page |
| Speklens | CNN-Transformer encoder pretrained on 222K molecules for spectral representation learning. | Code |
| Library | Description | Install |
|---|---|---|
| SpectraKit | Python toolkit for spectral preprocessing — baseline correction, smoothing, normalization, peak detection, multi-format I/O (JCAMP, SPC, OPUS, HDF5). Functional API, 699 tests, 2 core deps. | pip install pyspectrakit |
| ReactorTwin | Physics-constrained neural ODEs for chemical reactor digital twins. 7 hard constraint types, 5 Neural DE variants, 8 reactor types, full digital twin stack (EKF, fault detection, MPC, meta-learning). Production stable. | pip install reactor-twin |
| SpectraView | Interactive React component for vibrational spectroscopy — Canvas+SVG hybrid rendering, LTTB downsampling, 4 file parsers (JCAMP-DX, CSV, JSON, SPC), spectral processing (baseline, normalization, SG smoothing, derivatives), comparison tools, snap crosshair, minimap, undo/redo. 15 components, 9 hooks, 266 tests. | npm i spectraview |
| Simulation | What it does | Demo |
|---|---|---|
| Symmetry Explorer | 27 molecules, 15 point groups, character tables, selection rules, vibrational mode animations, point group flowchart | Launch |
| Normal Mode Explorer | 45 molecules, 583 normal modes — side-by-side comparison, superposition, Boltzmann panel, sonification | Launch |
| Orbital Architect | Gamified quantum chemistry — build atoms by placing electrons following Aufbau, Pauli, and Hund's rules. 36 levels. | Launch |
| VibeScope | Real-time 3D molecular vibration visualization with IR/Raman spectrum overlay | Launch |
| Spectrum-to-Structure | Input a vibrational spectrum, watch an ML model predict the 3D molecular structure | Coming soon |
| SpectraView Demo | Interactive IR spectrum viewer with peak detection, region highlighting, and multi-spectrum overlay | Storybook · GitHub |
- Choosing Physics Constraints for Neural ODE Reactor Surrogates — empirical comparison of hard projection vs soft penalty constraints across 3 benchmarks, 7 conditions, 3 seeds. ML4PS @ NeurIPS 2026.
- Can One Hear the Shape of a Molecule? Group-Theoretic Identifiability and Modal Complementarity in Vibrational Spectroscopy — first formal identifiability theory for the spectral inverse problem
- Hybrid State-Space Attention for Multi-Task Vibrational Spectroscopy — the Spektron architecture: SSMs + transformers + MoE for spectral analysis
Blog — Latent Chemistry
Technical writing on deep learning for chemical sciences. Each post includes interactive visualizations and equations.
| Post | Topic |
|---|---|
| VIB for Spectral Disentanglement | Splitting chemistry from instrument with gradient reversal and beta annealing |
| Masked Pretraining for Scientific Spectra | BERT-style masking for continuous 1D signals, identity collapse pitfall |
| Optimal Transport for Spectral Matching | Sinkhorn OT as a geometry-aware loss function |
| The Spectral Inverse Problem | From Wilson GF equation to symmetry-aware deep learning |
| SpectraView: Canvas-First Visualization | Building a high-performance spectral viewer in React |
| Spectral Identifiability Theory | Group-theoretic constraints on structure determination |
| SpectraKit: A Functional API | Designing a minimal spectral preprocessing library |
| State Space Models for Spectroscopy | Why sequence models beat CNNs for spectra |
| Neural ODEs for Reactor Modeling | Physics-constrained continuous-time dynamics |
| Why Spectra Are Harder Than Images | Sharp peaks, instrument drift, no ImageNet |
Python PyTorch CUDA TypeScript React Next.js Astro Three.js Tailwind CSS RDKit Docker D-LinOSS/SSMs Transformers Optimal Transport
Vibrational Spectroscopy · Spectral Inverse Problem · Self-Supervised Pretraining · Molecular Representation Learning · Information-Theoretic Chemistry · Physics-Informed ML · Calibration Transfer · Group Theory in Spectroscopy · Optimal Transport