SMT Explainability 🧠

Enhancing SMT with Explainable AI for Surrogate Models

A compact extension to the Surrogate Modeling Toolbox (SMT) that brings Explainable AI (XAI) techniques to surrogate models used in engineering: interpretation, visualization, and statistically-valid uncertainty.

---

About

SMT Explainability is an extension to the SMT (Surrogate Modeling Toolbox). It provides a collection of model-agnostic XAI methods tailored to surrogate models (Kriging, neural networks, etc.), with special care for mixed continuous/categorical inputs common in engineering design.

Key Features

• Model-agnostic interpretation — Works with a wide range of surrogate models to reveal how predictions are produced.
• SHAP (Shapley Additive Explanations) — Local and global explanations using Shapley values.
• PDP & ICE plots — Partial Dependence and Individual Conditional Expectation visualizations for marginal effects and heterogeneity.
• Mixed-variable support — Handles continuous and categorical (including mixed-integer) variables.
• Uncertainty quantification — Split conformal prediction for robust, distribution-free prediction intervals.
• Global sensitivity analysis — Sobol indices for variance-based sensitivity and corresponding visual reports.

Why It Matters

Surrogate models replace expensive simulations in engineering workflows, but they can act as opaque black boxes. SMT Explainability helps engineers:

• Build trust by making model decisions transparent.
• Discover insight such as non-linearities, interactions, and variable importance.
• Support decisions with quantified contributions (why the model predicted X for a given input).

---

Installation

pip install smt-explainability

Or install from source:

git clone https://github.com/your-org/smt-explainability.git
cd smt-explainability
pip install -e .

Requirements

See requirements.txt for full details. Minimal requirements:

• smt (Surrogate Modeling Toolbox)
• numpy
• matplotlib
• scikit-learn

---

Reference / Citation

If you use this package in research, please cite:

@inproceedings{robani2025smtex,
  title={SMT-EX: An Explainable Surrogate Modeling Toolbox for Mixed-Variables Design Exploration},
  author={Robani, Mohammad Daffa and Saves, Paul and Palar, Pramudita Satria and Zuhal, Lavi Rizki and Morlier, Joseph},
  booktitle={AIAA SCITECH 2025 Forum},
  year={2025},
  doi={10.2514/6.2025-0777}
}

A short description of the validation cases (wing weight prediction, beam bending) and methodology is contained in the paper.

---