SOUMIK GHOSH BLACKNIGHT-cloud12

Hey, I'm Soumik Ghosh 🧪✈️📈

I am a Scientific Machine Learning (SciML) and Computational Engineering practitioner working at the intersection of fluid dynamics, physics-informed deep learning, and quantitative modeling.

My work focuses on building architectures that do not treat physical systems as black-box interpolators, but instead respect the underlying conservation laws (Navier-Stokes mass, momentum, and energy conservation, frame indifference, and thermodynamics).

🌟 Featured Engineering & Research Projects

✈️ NeuroFoil: Industrial SciML Aerodynamic Platform

End-to-end automated OpenFOAM-to-SciML surrogate and design platform

Predicts aerodynamic lift ($C_L$), drag ($C_D$), moment ($C_m$), and continuous 161-point surface pressure fields $C_p(x)$ with sub-millisecond latency ($R^2 = 0.992$).
Coupled Aero-Structural Wingbox Mechanics: Sizing front/rear spar web depths, inter-spar fuel tank volume, and cross-sectional bending moment of inertia ($I_{xx}$).
Epistemic Uncertainty Quantification (UQ): Mahalanobis distance OOD detection against the training manifold covariance matrix with 95% conformal prediction bands.
MDAO SLSQP Optimizer: Constrained inverse design discovering optimal profiles under combined aerodynamic and structural bounds.

Python OpenFOAM v2606 PyTorch SciPy Streamlit MDAO WSL2

🌪️ tbnnFoam: Data-Driven Turbulence Closures in OpenFOAM

Galilean-invariant Tensor Basis Neural Networks for RANS Reynolds stress anisotropy

Couples Pope's tensor basis representation with OpenFOAM's $k$-$\omega$ SST and RANS framework to predict the anisotropic Reynolds stress tensor discrepancy $\tau_{ij}^\Delta$.
Enforces material frame indifference (objectivity) and Galilean invariance under spatial rotations, addressing the failure of linear eddy-viscosity Boussinesq closures under strong adverse pressure gradients.
Integrated into OpenFOAM C++ solver loops via LibTorch runtime bindings.

C++ OpenFOAM LibTorch PyTorch Turbulence Modeling Tensor Algebra

🎯 TO-PINN-Aerodynamics: Unsupervised Aerodynamic Topology Optimization

Topology optimization directly governed by Navier-Stokes PDE loss constraints

Formulates aerodynamic shape and topology optimization as an unsupervised continuous field learning problem without relying on costly discrete adjoint CFD solvers.
Embeds density penalty methods (SIMP) directly within Physics-Informed Neural Network (PINN) loss formulations.
Discovers minimum drag topologies under volume fraction constraints.

Python PyTorch PINNs Automatic Differentiation Topology Optimization

📐 NTK-PTC: Spectral Bias Control in PINNs

Neural Tangent Kernel (NTK) analysis & Phase Transition Control for stiff PDEs

Investigates the spectral bias and eigenvalue decay of the empirical Neural Tangent Kernel in Physics-Informed Neural Networks applied to convective and stiff boundary layer problems.
Develops dynamic loss balancing algorithms based on NTK trace alignment to prevent optimization collapse during high-frequency physical transitions.

Python PyTorch NTK Theory Stiff PDEs Numerical Analysis

🌊 AeroPINO: Physics-Informed Neural Operator for Aerodynamics

Mesh-independent flow field mapping using Fourier Neural Operators (FNO)

Combines data-driven Fourier Neural Operators with physics-informed PDE residual penalties (PINO) to learn continuous velocity and pressure fields over arbitrary body geometries.
Achieves zero-shot super-resolution and transferability across grid resolutions.

Python PyTorch Fourier Neural Operators (FNO) PINO SciML

📈 Real Estate LOB SSM: Continuous-Time Market Depth Simulator

Deep State Space Model (SSM) for Limit Order Book counterfactual simulations

Uses structured state space sequences (S4/Mamba architecture) to model continuous-time order flow arrival, microstructural liquidity, and bid-ask spread dynamics.
Nested with an interactive Next.js cockpit for counterfactual liquidity shock testing.

Python PyTorch State Space Models (SSM) Next.js TypeScript FastAPI

🔬 Core Technical & Research Focus

Research Domain	Core Focus	Methodologies & Tooling
Scientific Machine Learning (SciML)	Physics-Informed Neural Networks, Operator Learning, NTK Spectral Bias	PINNs, DeepONet, Fourier Neural Operators (FNO), PINO, JAX
CFD & Turbulence Modeling	Data-driven closures, structured boundary layer meshing, Reynolds stress	OpenFOAM (v2206/v2606), $k$-$\omega$ SST, TBNN, $y^+ \le 1$ C-Grids
Aerodynamic Design & MDAO	Parametric shape exploration, aero-structural wingbox sizing, inverse design	Kulfan CST, SLSQP, Area Moment of Inertia $I_{xx}$, Beam Sizing
Uncertainty Quantification (UQ)	Conformal prediction intervals, out-of-distribution detection in physics	Mahalanobis distance, Deep Ensembles, Covariance manifolds
Foundation Models & Quant Systems	Parameter-efficient fine-tuning, continuous-time state space models	Gemma post-training (PEFT/LoRA), S4 / Mamba, Stochastic Calculus

⚡ Technical Stack

💻 Languages & Systems

🤖 Machine Learning & SciML

🔬 Computational Fluid Dynamics & Engineering

⚙️ Production & Application Deployment

📊 GitHub Analytics & Activity

📚 Foundational Literature & Influences

"The purpose of computing is insight, not numbers." — Richard W. Hamming

Field	Seminal Reference	Engineering Significance
PINNs	Raissi et al. (2019) — Physics-Informed Neural Networks	Core paradigm of embedding PDE residuals into computational loss functions.
Neural Operators	Lu et al. (2021) — Learning Nonlinear Operators via DeepONet	Universal approximation of nonlinear continuous operators across function spaces.
Spectral Solvers	Li et al. (2020) — Fourier Neural Operator for Parametric PDEs	Mesh-independent parameterization of integral kernels directly in Fourier domain.
Turbulence ML	Ling, Kurzawski & Templeton (2016) — Reynolds Stress Anisotropy (TBNN)	Embedding Galilean and rotational tensor invariance into neural turbulence closures.
Data-Driven Dynamics	Brunton & Kutz — Data-Driven Science and Engineering	Synthesis of system identification (SINDy), Koopman operator theory, and control.
Quantitative Calculus	Paul Wilmott — On Quantitative Finance	Stochastic differential equations, Fokker-Planck PDEs, and martingale pricing.

Open to collaborations in Scientific Machine Learning, Physics-Informed AI, and Aerodynamic Engineering.

Navigation Menu