dhrub chatterjee dhrubchatterjee16-lang

⚛️ DHRUB CHATTERJEE

Chemical Engineer · Scientific Machine Learning · Autonomous Discovery

🧬 Executive Profile

I engineer at the intersection of Chemical Process Engineering and Foundational Scientific AI. Rather than treating machine learning as an unconstrained black-box, my research and architectures enforce physical conservation laws, thermodynamic stability criteria, and geometric symmetries directly into computational graphs.

      Classical Thermodynamics             Geometric Deep Learning              Closed-Loop Automation
 ┌───────────────────────────────┐     ┌───────────────────────────────┐     ┌──────────────────────────────┐
 │ • Helmholtz Free Energy (ICNN)│     │ • SE(3) Equivariant GNNs      │     │ • Multi-Fidelity Bayesian Opt│
 │ • CAPE-OPEN / Peng-Robinson   │ ──> │ • Molecular Force Fields      │ ──> │ • Opentrons Robotic Control  │
 │ • ISA-95 Industrial Twins     │     │ • Polymer Topological Graphs  │     │ • Self-Driving Catalyst Labs │
 └───────────────────────────────┘     └───────────────────────────────┘     └──────────────────────────────┘

🏛️ Research & Architecture Taxonomy

flowchart LR
    classDef core fill:#1e1e2e,stroke:#6c5ce7,stroke-width:2px,color:#fff;
    classDef flag fill:#0f3460,stroke:#00fff5,stroke-width:2px,color:#fff;
    classDef tool fill:#16213e,stroke:#e94560,stroke-width:1px,color:#fff;

    Chem["<b>Domain Foundations</b><br/>Thermodynamics & Kinetics"]:::core

    Chem --> P1["<b>1. Thermodynamically Consistent ML</b>"]:::core
    Chem --> P2["<b>2. Geometric & Materials AI</b>"]:::core
    Chem --> P3["<b>3. Autonomous Discovery & SDLs</b>"]:::core
    Chem --> P4["<b>4. Domain-Adapted Scientific LLMs</b>"]:::core

    P1 --> F1["<b>pinn-thermo</b><br/>ICNN Helmholtz Engine"]:::flag
    P1 --> F2["<b>DeepRefine</b><br/>Level 3 RTO & ISA-95"]:::flag
    P1 --> F3["<b>corrosense-AI</b><br/>NumPy Nernst-Planck"]:::flag

    P2 --> F4["<b>chem-scientist-ai</b><br/>SE(3) Continuous SchNet"]:::flag
    P2 --> F5["<b>molgenie</b><br/>Grammar Retrosynthesis"]:::flag
    P2 --> F6["<b>PolymerGenesis</b><br/>D-MPNN Polymer Physics"]:::flag

    P3 --> F7["<b>auto-catalyst-2.0</b><br/>BoTorch + Opentrons SDL"]:::flag

    P4 --> T1["<b>dapt-llm-domain-adapter</b>"]:::tool
    P4 --> T2["<b>strict-structured-json-extractor</b>"]:::tool

🚀 Flagship Deployments

⚡ DeepRefine

Industrial Level 3 Supervisory RTO & Digital Twin Testbed

Bridges ISA-95 / IEC 62443 manufacturing automation with Physics-Informed Neural Networks.
Real-time optimization (RTO) resolving stiff differential-algebraic equations (DAEs) under <100ms cycle constraints.
Built-in API 670 machinery vibration diagnostics and MPC supervisory control.

PyTorch PINNs ISA-95 Process Control MPC

🧪 auto-catalyst-2.0

Autonomous Self-Driving Lab (SDL) for FCC Catalysts

Closed-loop active learning orchestrating multi-fidelity Bayesian optimization via BoTorch & Ax.
Constrained multi-objective acquisition ($q\text{NEHVI}$) balancing activity, selectivity, and catalyst coking kinetics.
Automated fluidics control scripts designed for Opentrons OT-2 liquid-handling robotics.

BoTorch Active Learning Opentrons Bayesian Opt

📐 pinn-thermo

ICNN Helmholtz Free Energy & CAPE-OPEN Engine

Formulates thermodynamically consistent surrogates: convex in volume (mechanical stability: $\kappa_T > 0$) and concave in temperature ($C_v > 0$).
Enforces $C^2$-continuity using GELU/SiLU activations for exact autograd state evaluation ($P = -\partial A/\partial V$).
Validated against Peng-Robinson (PR-EOS) and PC-SAFT baselines.

ICNN Thermodynamics CAPE-OPEN SciML

🔮 chem-scientist-ai

SE(3)-Equivariant Continuous-Filter Molecular Intelligence

Geometric Graph Neural Network enforcing translational and rotational equivariance over 3D Cartesian coordinates.
Predicts invariant scalar potentials (energy) and equivariant vector force fields ($\mathbf{F} = -\nabla_{\mathbf{r}} E$).
Evaluated against standard atomistic benchmarks (QM9, OC20).

Geometric DL SE(3) Equivariance SchNet PyG

⚙️ corrosense-AI

Electrochemical Flow Corrosion Intelligence from Scratch

Implemented end-to-end in pure NumPy with zero framework abstractions (custom autograd, Adam, BatchNorm, LeakyReLU).
Models multiphase Nernst-Planck ionic transport and electrochemical Butler-Volmer boundary kinetics.

Pure NumPy Autograd from Scratch Electrochemistry

🧬 molgenie

Grammar-Guaranteed Generative AI & Retrosynthesis

Constrained molecular generation engine guaranteeing 100% valence validity via context-free SMILES grammars.
Evaluates retrosynthetic accessibility (SAScore / SCScore) and reaction template compatibility.

Generative AI Retrosynthesis RDKit Cheminformatics