A Look into How Machine Learning is Reshaping Engineering Models: the Rise of Analysis Paralysis, Optimal yet Infeasible Solutions, and the Inevitable Rashomon Paradox

Technical Summary: A Look into How Machine Learning is Reshaping Engineering Models

Problem Statement
The paper addresses a philosophical and practical tension within civil engineering: the widespread acceptance of empirically derived codal provisions (e.g., Manning equation, shear strength equations) versus the skepticism surrounding Machine Learning (ML) models. While both rely on statistical foundations, traditional empirical formulas are trusted due to decades of validation, transparency in variable interaction, and alignment with engineering judgment. In contrast, ML models are often viewed as "black boxes" that lack physical interpretability, challenging the traditional engineering mindset which prioritizes understanding mechanisms over pure predictive accuracy. The paper argues that this hesitation stems from a disconnect between the digitalized nature of ML and the familiar mathematical forms of empirical equations, as well as a lack of standardization and historical validation for ML in engineering practice.

Methodology
The study employs a multi-faceted approach combining philosophical discussion with empirical case studies to demonstrate how ML can be integrated into engineering reasoning via deduction, induction, and abduction.

1. Model Hierarchy Analysis: The paper categorizes engineering models from physics-based (e.g., Finite Element Analysis) to purely data-driven, mapping them to ML equivalents such as Physics-Informed Neural Networks (PINNs) and causal discovery models.
2. Case Studies in Reasoning:
   • Induction: Using the PySR (Symbolic Regression) library on a concrete compressive strength dataset (1,030 samples), the study derives new mathematical expressions. These are compared against Abram's Law and traditional ML algorithms (Random Forest, XGBoost, etc.) to evaluate predictive accuracy and feature importance.
   • Abduction: A Random Forest model is trained on fire resistance data for Reinforced Concrete (RC) columns (140 tests) to identify structural configurations that meet a 120-minute fire resistance target. The study utilizes SHAP (SHapley Additive exPlanations) values for interpretability and employs unsupervised clustering to generate hypotheses for high-performing configurations.
   • Deduction: The paper discusses the limitations of Euler's Buckling formula (idealized assumptions) and proposes how ML can calibrate Finite Element models or use Bayesian approaches to account for real-world imperfections, though a specific new case study for deduction is omitted to avoid redundancy with existing literature.
3. Paradox Investigation: The study identifies and analyzes three specific paradoxes arising from ML adoption:
   • Analysis Paralysis: High accuracy at the cost of physical understanding.
   • Optimal yet Infeasible Solutions: Optimization yielding designs that violate constructability or standardization.
   • The Rashomon Paradox: Contradictory explanations from different interpretability methods (SHAP vs. LIME) that may conflict with physical laws.

Key Results

• Induction (Concrete Strength): Symbolic regression (PySR) generated expressions for concrete compressive strength that significantly outperformed Abram's Law (R² = 0.926 vs. 0.724) when using water-cement ratio alone. When incorporating all mixture proportions, ML models (Gradient Boosting, LightGBM, XGBoost) achieved R² values up to 0.9477, outperforming traditional empirical extensions. Feature importance analysis confirmed that ML identified physically relevant variables (e.g., water-cement ratio, age, cement content).
• Abduction (Fire Resistance): The Random Forest model achieved an R² of 0.748 on the testing set for predicting RC column fire resistance. SHAP analysis successfully identified configurations meeting the 120-minute target, revealing that column height and eccentricity reduce fire resistance, while cover thickness and width increase it. Unsupervised clustering successfully grouped data into five distinct clusters based on loading and material properties, allowing for the generation of new, plausible configurations for a 240-minute target.
• Paradoxes Identified:
  • Analysis Paralysis: A symbolic regression model for Concrete-Filled Steel Tube (CFST) axial capacity outperformed AISC 360, Eurocode 4, and AS 2327 provisions in terms of MAE and RMSE. However, the resulting expression lacked clear physical justification, illustrating the trade-off between accuracy and interpretability.
  • Infeasible Solutions: An optimization study for CFST columns using the SPINEX algorithm produced a theoretically optimal solution (380 mm diameter) that satisfied fire resistance and volume constraints but was infeasible due to the non-existence of a standard 380 mm steel section.
  • Rashomon Paradox: Comparing SHAP and LIME on the RC column model revealed conflicting feature attributions. For instance, one method suggested higher yield strength improved fire resistance, while the other suggested it reduced it, contradicting established thermomechanical principles despite the model's high statistical accuracy.

Significance and Claims
The paper claims that while ML offers superior predictive capabilities and the ability to handle complex, non-linear relationships in structural engineering, its integration requires a rethinking of engineering epistemology and education. The author argues that ML should not replace traditional methods but serve as an extension of the engineer's toolkit.

The significance of the work lies in its identification of three critical paradoxes that must be resolved for ML to be safely adopted:

1. Bridging the Interpretability Gap: Engineers must move beyond "black box" acceptance by using hybrid approaches (e.g., Physics-Informed Neural Networks) and validation frameworks that ensure ML predictions align with physical mechanisms.
2. Ensuring Constructability: Optimization algorithms must incorporate discrete constraints and knowledge-based systems to prevent the generation of theoretically optimal but practically unviable designs.
3. Resolving Explainability Conflicts: The paper advocates for a multi-method approach to explainability, where results from tools like SHAP and LIME are synthesized and filtered through a layer of physical verification to ensure feature attributions do not contradict fundamental engineering laws.

Ultimately, the paper concludes that the future of engineering education and practice must harmonize traditional principles with ML, addressing the diminishing role of human intuition and the need for standardized, interpretable, and ethically sound ML frameworks in the profession.