[Paper Review] Explainable AI for Engineering Design: A Unified Approach of Systems Engineering and Component- Based Deep Learning Demonstrated by Energy- Efficient Building Design

Philipp Geyer, Manav Mahan Singh+1 more|arXiv (Cornell University)|Aug 30, 2021

Probabilistic and Robust Engineering Design33 references4 citations

TL;DR

This paper proposes a component-based deep learning framework that unifies systems engineering with explainable AI for engineering design, using interpretable interface activations to enhance model transparency and reusability. In energy-efficient building design, it achieves superior generalization (R² = 0.94 vs. 0.71) and high alignment with physical simulations (R² = 0.92–0.99 for envelopes, 0.78–0.93 for zones), demonstrating improved explainability through sensitivity analysis and decision trees.

ABSTRACT

Data-driven models created by machine learning, gain in importance in all fields of design and engineering. They, have high potential to assist decision-makers in creating novel, artefacts with better performance and sustainability. However,, limited generalization and the black-box nature of these models, lead to limited explainability and reusability. To overcome this, situation, we propose a component-based approach to create, partial component models by machine learning (ML). This, component-based approach aligns deep learning with systems, engineering (SE). The key contribution of the component-based, method is that activations at interfaces between the components, are interpretable engineering quantities. In this way, the, hierarchical component system forms a deep neural network, (DNN) that a priori integrates information for engineering, explainability. The, approach adapts the model structure to engineering methods of, systems engineering and to domain knowledge. We examine the, performance of the approach by the field of energy-efficient, building design: First, we observed better generalization of the, component-based method by analyzing prediction accuracy, outside the training data. Especially for representative designs, different in structure, we observe a much higher accuracy, (R2 = 0.94) compared to conventional monolithic methods, (R2 = 0.71). Second, we illustrate explainability by exemplary, demonstrating how sensitivity information from SE and rules, from low-depth decision trees serve engineering. Third, we, evaluate explainability by qualitative and quantitative methods, demonstrating the matching of preliminary knowledge and data-driven, derived strategies and show correctness of activations at, component interfaces compared to white-box simulation results, (envelope components: R2 = 0.92..0.99; zones: R2 = 0.78..0.93).

Motivation & Objective

• To address the black-box nature and limited reusability of conventional machine learning models in engineering design.
• To improve model generalization beyond training data distributions, especially for structurally diverse designs.
• To integrate domain-specific knowledge from systems engineering and physical simulation into deep learning architectures.
• To enable explainability through interpretable interface activations and rule-based interpretation methods.
• To validate the approach using real-world energy-efficient building design with quantitative and qualitative explainability metrics.

Proposed method

• The method decomposes the design problem into hierarchical components, each modeled as a separate deep neural network (DNN) with domain-specific inputs and outputs.
• Interface activations between components are constrained to represent interpretable engineering quantities (e.g., thermal transmittance, zone temperature).
• The overall system forms a deep neural network where component connections reflect physical and systems engineering relationships.
• Model structure is informed by systems engineering principles, including modularity, abstraction, and interface standardization.
• Explainability is enhanced via sensitivity analysis and low-depth decision trees trained on activation patterns.
• Performance is validated by comparing predictions against white-box simulation results using R² correlation metrics.

Experimental results

Research questions

• RQ1 Can a component-based deep learning architecture improve generalization performance on out-of-distribution building designs compared to monolithic models?
• RQ2 To what extent do interface activations in the component model align with physically meaningful engineering quantities?
• RQ3 How effectively can sensitivity analysis and decision trees extract interpretable insights from the model’s internal representations?
• RQ4 How well does the component-based model match the results of high-fidelity white-box simulations for building envelope and zone components?
• RQ5 Can the model’s predictions be trusted and validated using domain knowledge and simulation benchmarks?

Key findings

• The component-based model achieved R² = 0.94 in predicting energy performance for structurally different buildings, significantly outperforming monolithic models (R² = 0.71).
• Interface activations for envelope components showed strong correlation with simulation results (R² = 0.92–0.99), confirming physical interpretability.
• Zone-level predictions also demonstrated high fidelity, with R² = 0.78–0.93 compared to simulation outputs.
• Sensitivity analysis and decision trees successfully extracted interpretable rules from model activations, aligning with engineering intuition.
• The model exhibited superior generalization, particularly on novel and structurally diverse design configurations.
• The integration of systems engineering principles into deep learning architecture enabled both high performance and explainability from the outset.