Jishnu Vinayak Gopi

A fully integrated Python-driven virtual laboratory that connects statistical mesostructure generation, thermo-mechanical finite element analysis, and simulation-based performance evaluation for MgO–C refractory composites. The framework enables systematic parametric studies of thermal shock resistance without physical specimens, bridging the gap between mesostructure design and macroscopic material behaviour.

• Laguerre–Voronoi tessellation-based mesostructure generator with statistical phase size distributions for realistic grain morphology.
• Coupled transient thermal–structural FEM in Abaqus with temperature-dependent constitutive data for MgO and carbon phases.
• Fracture-metric parameter for quantitative thermal shock resistance assessment from crack length and density fields.
• Python automation layer for mesh generation, job submission, field output extraction, and result visualisation.

A Fortran-based Abaqus User Element (UEL) implementing a fully coupled thermo-mechanical cohesive zone model for simulating debonding at MgO aggregate–matrix interfaces driven by thermal expansion mismatch. The formulation accounts for temperature-dependent fracture properties and consistent tangent stiffness for quadratic convergence.

• Thermodynamically consistent cohesive traction–separation law with thermal expansion mismatch as a damage driver.
• Four-stage interface thermal conductance model governing heat transfer normal to the cohesive zone: intact interface (Kapitza thermal resistance), partial separation (Kapitza + gas + radiation), full separation (gas + radiation), and re-contact (gas + radiation + surface contact conductance).
• Consistent algorithmic tangent for coupled thermal and mechanical degrees of freedom ensuring robust Newton–Raphson convergence.
• Validated against analytical solutions and benchmark problems for interface debonding under thermal loading.
• Integrated into the virtual laboratory mesh pipeline via Python-automated interface element insertion.

A two-scale finite element homogenisation framework (FE²) for woven fibre-reinforced composites, coupling macroscopic structural responses to microscale representative volume element (RVE) simulations. The implementation follows the Hill–Mandel macrohomogeneity condition and supports periodic, Dirichlet, and Neumann boundary conditions on the RVE.

• Macroscale–microscale coupling via consistent tangent moduli derived from RVE condensation.
• Python-driven woven RVE generator producing geometry and mesh for parametric weave architectures.
• MPI-parallelised C++ solver for RVE computations, reducing wall-clock time for large parametric sweeps.
• Validated effective elastic properties against analytical bounds (Voigt, Reuss, Hashin–Shtrikman).

A physics-informed data-driven pipeline for identifying nonlinear material hardening parameters from small punch test (SPT) force–displacement data. Abaqus FEM simulations were used to build a structured training dataset, and a feedforward neural network (FFNET) was trained to perform the inverse mapping from punch curve features to constitutive parameters.

• Automated Abaqus Python scripting for systematic generation of labelled simulation datasets across parameter space.
• Feature extraction from force–displacement curves (energy, slope, characteristic points) as neural network inputs.
• Feedforward network trained on simulation data; validated against held-out FEM cases and experimental SPT results.
• Demonstrated significant reduction in identification time compared to iterative optimisation approaches.

A parallel image processing program implementing Gaussian blur, sharpening, and Laplacian edge detection on large grayscale PGM images using C++ and MPI. The design decomposes the image into row-wise batches distributed across ranks, exchanges halo rows via non-blocking point-to-point communication with periodic rank topology, and collects results with MPI_Gather for PGM file output. Benchmarked on the TU Freiberg cluster up to 36 processes across three nodes.

• Distributed file I/O: each rank independently opens the PGM file, skips rows outside its assigned batch using a pixel-read loop, and stores only its own batchSize × numberCol slice — avoiding any broadcast of the full image.
• Non-blocking halo exchange: MPI_Isend / MPI_Irecv with periodic rank topology (rank 0 ↔ rank size−1) for correct boundary treatment at the top and bottom of the domain; followed by MPI_Barrier before convolution.
• Three convolution kernels: Gaussian blur (3×3, weights stored as floats, 1/16 normalisation), sharpen, and Laplacian edge detection — all with periodic column-wise boundary handling and [0, 255] pixel clamping.
• Chained pipeline for Task 3: performConvolution() is parameterised by flag (B/S/E) and iteration count; Task 3 chains 5× blur → 1× sharpen → 1× edge detection through intermediate PGM files, each stage collecting via MPI_Gather on rank 0.
• Strong scalability: wall-time for the full Task 3 pipeline (5× blur → sharpen → edge detection) on an 8162×8162 image reduced from 133.6 s (1 process) to 39.9 s (36 processes), a ~3.3× speedup; gains taper beyond 12 processes as communication and I/O overhead dominate.