When applying machine learning to physical systems—whether predicting the failure of a pipeline, the deformation of a material, or the energy consumption of a building—standard data-driven models face a fundamental limitation: they are physically blind. A vanilla neural network treats physical variables as arbitrary numbers, completely unaware of the laws of thermodynamics, fluid dynamics, or conservation of energy.
To fix this, the dominant paradigm has long been Physics-Informed Neural Networks (PINNs), which enforce physical laws by adding regularization terms to the loss function. However, optimizing these multi-objective loss functions is notoriously difficult, often leading to competing gradients and training instability.
An alternative, highly effective direction has emerged: Physics-Informed Feature Augmentation. Instead of forcing a model to discover physical laws through training penalties, researchers inject physical domain knowledge directly into the input feature space. By giving the model a physically grounded starting point, this approach linearizes complex relationships, constrains the search space, and drastically reduces the need for massive architectures.
Shifting Physics from the Loss to the Input Layer
The core intuition behind feature augmentation is simple: it is much easier for a model to learn a pattern if the underlying physical dynamics are already explicitly represented in its inputs.
As Majd Al Aawar et al. demonstrate in Dynamics-Based Feature Augmentation of Graph Neural Networks for Variant Emergence Prediction, integrating domain-specific dynamics directly as input features is significantly more effective than using them as loss function constraints. In their work, providing the dynamics explicitly linearizes complex exponential problems, bypassing the optimization difficulties typically associated with physics-informed loss terms.
Similarly, Shaghayegh Fazliani shows in Enhancing Physics-Informed Neural Networks Through Feature Engineering that augmenting the input space with mathematical functions that mirror a problem's initial and boundary conditions allows even single-layer networks to satisfy physical constraints easily. Rather than relying on deep feature discovery to stumble upon the boundaries of a physical system, the network is handed the boundaries at the outset.
Transforming Raw Data into Physical Descriptors
One primary mechanism of physics-informed feature engineering is the transformation of raw, uninformative inputs into high-fidelity physical descriptors or analytical derivatives.
In materials science and chemistry, feeding raw chemical compositions into a model forces it to infer complex atomic interactions from scratch. To circumvent this, Leonardo Bertolucci Coelho et al. in Reviewing machine learning of corrosion prediction in a data-oriented perspective transform raw composition data into fundamental physical descriptors such as electronegativity and molar volume. This strategy allows the model to reflect the actual mechanisms of chemical degradation rather than relying on simple statistical correlations. Along the same lines, Alexandre Duval et al. in PhAST: Physics-Aware, Scalable, and Task-Specific GNNs for Accelerated Catalyst Design integrate periodic table group information and electronegativity directly into atom embeddings, significantly boosting the model's ability to generalize to unseen chemical compositions.
For dynamical systems characterized by rapid changes or discontinuities, raw positional or spatial data is rarely enough. In Physics-Informed Spatiotemporal Deep Learning for Emulating Coupled Dynamical Systems, Anishi Mehta et al. incorporate first and second-order partial derivatives of physical fields as additional input channels. This provides the network with explicit information about underlying gradients, enabling it to accurately emulate fracture propagation.
This approach also proves vital for out-of-distribution generalization. Wannes De Groote et al. in Physics-Based Neural Network Models for Prediction of Cam-Follower Dynamics Beyond Nominal Operations augment input features by pre-calculating analytical derivatives and physical relations from a system's governing equations. Because the network is grounded in these fundamental relations, it successfully predicts hazardous mechanical failures even when trained exclusively on safe, nominal operating data.
Integrating Low-Fidelity Physical Models as Auxiliary Inputs
Another powerful strategy is to use the outputs of simplified, deterministic physical models or classical empirical equations as auxiliary input features for a more complex machine learning architecture.
In building energy management, traditional physics models are computationally heavy but grounded in thermodynamics, while data-driven models are fast but physically blind. Leandro Von Krannichfeldt et al. in Integrating Physics-Based and Data-Driven Approaches for Probabilistic Building Energy Modeling resolve this by using the outputs of a simplified physical model as auxiliary input features for a deep learning model. This provides a physically grounded baseline that keeps predictions realistic, even when data is scarce.
This hybrid framing appears across diverse engineering disciplines:
- Geotechnical Engineering: Te Pei et al. in Closure to “Applying Knowledge-Guided Machine Learning to Slope Stability Prediction” integrate domain-specific metrics derived from physical equations directly into the input layer. These calculations provide a high-leverage signal that dominates the model's predictive logic, compensating for limited empirical datasets.
- Industrial Drilling: Mengyuan Xiong et al. in A rate of penetration (ROP) prediction method based on improved dung beetle optimization algorithm and BiLSTM-SA feed the results of classical empirical equations into a neural network as additional features, constraining the search space and providing a theoretically grounded starting point.
- Traffic Forecasting: Steven Wong et al. in Traffic Forecasting using Vehicle-to-Vehicle Communication incorporate the outputs of a deterministic physical traffic model as auxiliary inputs to a recurrent neural network, combining theoretical priors with real-time sensor data.
Why It Works: Search Space Constraint and Architecture Simplification
Physics-informed feature augmentation works because it fundamentally alters the learning landscape. Instead of forcing a network to map a high-dimensional raw input space to an output space, augmented features act as a mathematical guide, narrowing down the physically permissible solutions.
First, it acts as a powerful regularizer. Xingyuan Miao et al. demonstrate this in Residual strength prediction of corroded pipelines based on physics-informed machine learning and domain generalization by combining raw physical dimensions with non-dimensional ratios and mechanical relationships calculated from established formulas. This input-level constraint prevents the model from making physically impossible predictions and significantly enhances its ability to generalize to unseen scenarios.
Second, it shifts the engineering burden away from architectural complexity. Relying entirely on a network to extract features from raw data often requires massive, deep architectures that suffer from slow training times, as noted by Weiyu Kong et al. in Fault detection of integrated navigation system based on genetic algorithm optimization for deep belief network.
When physical properties are structured beforehand, simple architectures can outperform massive black-box models. Hyun-Myung Woo et al. show in Accelerating Optimal Experimental Design for Robust Synchronization of Uncertain Kuramoto Oscillator Model Using Machine Learning that sorting and extracting physically relevant features (like coupling strengths and frequency differences) allows simple neural networks to achieve high accuracy without needing complex neural differential equations.
This principle is reinforced by Parand Akbari et al. in MeltpoolNet: Melt pool Characteristic Prediction in Metal Additive Manufacturing Using Machine Learning, who find that incorporating domain-specific mathematical proxies into the input space improves model accuracy far more effectively than increasing the complexity or depth of the neural network architecture itself.
Ultimately, physics-informed feature augmentation proves that a dollar's worth of domain knowledge injected at the input layer is often worth a hundred dollars of architectural engineering and brute-force compute.