The manuscript tunes four regional stiffness parameters. This interactive extension assigns a separate Eᵢ to every cell in a 20 × 10 mesh. The medium has one connected layer of 3D elements. A contiguous 20-cell damaged area has a lower target stiffness than the surrounding material. Watch all 200 parameters evolve through swarm search and bounded local refinement toward this nonuniform stiffness field.
24 candidate fields explore 200 dimensions. Every particle carries a complete vector [E₁, E₂, …, E₂₀₀], and every coordinate is updated from its inertia, personal best, and swarm best.
The best swarm field initializes a projected limited-memory BFGS demonstration. All 200 parameters are refined together with gradient information, a line search, and bounds.
All 200 cell parameters are distinct optimization variables. Neighboring elements within the same region can have different initial values and follow different paths. The selected cell’s trace and values update as learning progresses. The illustrative damaged patch contains 20 cells (columns 9–13, rows 4–7); its target stiffness is adjustable from 40% to 80% of intact stiffness. Changing it restarts learning.
pᵢ = Eᵢ / 500,000; rᵢ = pᵢ − tᵢ
tᵢ = 1 intact; tᵢ = ρ damaged
J(p) = mean[rᵢ²] + 0.04 mean[(rᵢ − rⱼ)²]
0.4 ≤ pᵢ ≤ 1.6
The second term compares neighboring residual corrections, preserving the reference stiffness jump across the damage boundary. This is a synthetic parameter-fitting objective with a known target, not the manuscript’s FEM/measurement mismatch. No finite-element response, contraction, or inverse identification is solved in this preview.
| Region | Initial E | Reported final E |
|---|---|---|
| E₁ | 450,000 | 500,011.9 |
| E₂ | 400,000 | 500,082.7 |
| E₃ | 550,000 | 499,899.3 |
| E₄ | 580,000 | 500,029.5 |
Figure 4 reports a target of 500,000 for each of the four regional parameters. Those reported values are separate from the newly computed 200-element demonstration above. Parameter units are not specified here.
Based on Mehrdad Shafiei Dizaji’s supplied technical manuscript, A Hybrid Training Algorithm for Continuum Deep Learning Neuro-Skin Neural Network. The manuscript proposes PSO followed by L-BFGS-B. This browser extension runs PSO followed by projected L-BFGS as an educational approximation of bounded local refinement; it is not a full L-BFGS-B implementation or a reproduction of the paper’s numerical study. The element-wise extension and 20-cell damage scenario are educational demonstrations, not reported damage-identification results from the manuscript. The damaged region remains part of the connected mesh; its elements have reduced stiffness and are not removed.