Wrapping up WCCM-ECCOMAS 2026 in Munich, Germany

I was so happy to take part in the World Congress in Computational Mechanics & ECCOMAS Congress (WCCM-ECCOMAS) 2026 in Munich, Germany. A few highlights below.

---

Contributed Talk

I had the chance to present our latest work, “Data-Driven Classification of Structural Nonlinearities Using Interpretable Deep Learning on Time Series”, co-authored with Dr. Gianmarco Mengaldo. It was great to talk through some of the challenges we faced — building such a large dataset, working with neural networks, and tackling explainability — and to brainstorm ideas with fellow attendees.

…

---

A technical walkthrough of attribution methods and how to evaluate them — assuming some comfort with math, but no prior ML background required.

---

The Problem: Black Boxes

Imagine you go to a doctor and they say: “You need surgery.”

You ask: “Why?”

They say: “I don’t know. That’s just what the model predicted.”

That’s the black box problem. A neural network takes in data, executes hundreds of thousands of nonlinear transformations, and produces a confident output, but the path from input to output is, for practical purposes, opaque. Even the engineers who built the model often can’t tell you which parts of the input drove a specific prediction.

…

Internal Seminar Series — MathExLab

Interpretable AI for Nonlinear Structural Dynamics and a Benchmark for Nonlinear Mode Interaction

In this internal seminar, I presented my ongoing research on interpretable AI for nonlinear structural dynamics and benchmark design for nonlinear mode interaction.

---

Seminar Details

Speaker	Bayan Abusalameh
Role	PhD Student, MathExLab
Date	24 February 2026
Time	2:00 pm (SGT)

---

Abstract

I introduced a large controlled dataset for detecting nonlinearities in vibrating structures directly from raw time-series signals, together with a post-hoc interpretability pipeline based on Integrated Gradients, DeepLIFT, GradientSHAP, and DeepSHAP, as well as new quantitative metrics for testing attribution fidelity.

…

By Bayan Abusalameh

TL;DR.

Linear oscillators are simple and elegant: each mode evolves independently, superposition holds, and responses remain fully predictable. Once nonlinearities enter—whether clearance, Coulomb friction, cubic stiffness (hardening or softening), or quadratic damping—the story changes. Resonances shift, signals distort, and energy begins to leak and exchange in unexpected ways.

Our benchmark captures this transition by generating controlled SDOF simulations that span both linear and nonlinear regimes, injecting realistic noise, and labeling each sample automatically. On top of this, we train neural networks and evaluate interpretability maps to reveal not only what the model predicts, but also why.

…

Why superposition breaks, energy starts to slosh, and what to look for in data

TL;DR. In the linear regime, each mode is an independent damped oscillator—clean, decoupled, and predictable. As amplitudes grow (or when damping/forcing aren’t “nice”), cross terms re-couple the modal equations. Near internal resonance (e.g., 1:1, 2:1, 3:1 ratios), those cross terms become near-resonant drives, and you see energy exchange, sidebands, and new frequencies. That’s “mode interaction.” Our benchmark builds controlled simulations that span both regimes and labels interaction strength automatically.

…