Standard neural networks treat all inputs as arbitrary scalars. When a model is trained to predict physical phenomena—such as fluid velocity, plasma stability, or structural stress—it typically receives raw measurements representing physical quantities (e.g., $10.5\text{ m/s}$ or $0.001\text{ Pa}\cdot\text{s}$). Without explicit physical guidance, the network must learn two things simultaneously: the arbitrary scaling of the chosen unit system and the actual underlying physical laws.
This lack of dimensional awareness creates severe overfitting to the training scale and catastrophic failure when encountering out-of-distribution (OOD) physical parameters. To bypass this limitation, a powerful paradigm has emerged: dimensionless feature representation. By transforming raw physical inputs into scale-invariant, unitless quantities before feeding them into machine learning architectures, models can capture universal physical relationships rather than unit-specific magnitudes.
The Core Mechanism: Dimensional Compression via Buckingham $\Pi$
How do we systematically strip units from data without losing physical information? The foundational mathematical tool is the Buckingham $\Pi$ theorem. As discussed by Abhishek Chandra et al. in ROLE OF PHYSICS IN PHYSICS-INFORMED MACHINE LEARNING, any physically consistent relationship involving $n$ variables with $m$ fundamental dimensions can be mapped to a lower-dimensional space of $n-m$ independent dimensionless parameters, denoted as $\Pi$ groups.
Consider fluid flow through a pipe fitting. Instead of feeding a model separate features for velocity, pipe diameter, fluid density, and viscosity, these variables are combined into a single dimensionless number: the Reynolds number ($Re = \rho v L / \mu$). As shown by Niklas Jarmatz et al. in Development of a soft sensor for fouling prediction in pipe fittings using the example of particulate deposition from suspension flow, this transformation compresses the input space without losing any governing physics. The machine learning model is freed from learning redundant scaling laws, allowing it to maintain high predictive accuracy while drastically simplifying the learning problem.
Why It Works: Similarity-Equivariance and Lab-to-Field Transfer
The fundamental reason dimensionless features work so well is the principle of physical similarity. In physics, two systems are dynamically similar if they share the same dimensionless numbers, regardless of their absolute physical size.
This property enables seamless scale-invariant generalization, allowing models to perform lab-to-field transfer. For instance, Evgenii Kanin et al. in A predictive model for steady-state multiphase pipe flow: Machine learning on lab data demonstrated that a machine learning model trained purely on small-scale laboratory data can accurately predict multiphase flow behaviors in massive industrial oil fields. By using dimensionless parameters like the Reynolds and Froude numbers, the model becomes invariant to changes in pipe diameter or fluid properties.
The same mechanism explains why Xiaomeng Dong et al. in An Evaluation of the Data-Driven Model for Bubble Maximum Diameter in Subcooled Boiling Flow Using Artificial Neural Networks achieved superior generalization across varying experimental configurations, and why Konstantia Nathanael et al. in Optimization of microfluidic synthesis of silver nanoparticles: a generic approach using machine learning successfully transferred chemical synthesis models across entirely different reactor geometries. The models do not memorize specific sensor values; they track the invariant ratios that dictate the underlying physical regimes.
Physics-Informed Normalization vs. Statistical Scaling
Traditional machine learning pipelines rely on statistical normalization (such as z-score or min-max scaling) to stabilize gradients. However, statistical scaling treats each variable independently, destroying the mathematical symmetries and ratios inherent in physical laws.
In contrast, physics-informed normalization scales data using primary physical constants or scalars from the system itself. This insight dates back to early neural network applications: Christopher M. Bishop in Real-Time Control of a Tokamak Plasma Using Neural Networks showed that normalizing plasma diagnostic signals by the total current incorporates linear invariance directly into the representation, preventing the network from needing to learn redundant scaling laws.
Modern deep learning architectures embed this directly into latent spaces and neural operators:
- Dimensionless Latent Spaces: In DimOL: Dimensional Awareness as A New 'Dimension' in Operator Learning, Yichen Song et al. designed neural operators that process partial differential equations (PDEs) within a dimensionless latent space, rendering them robust against out-of-distribution shifts in physical constants like viscosity or pressure.
- Geometric and Spatial Scaling: The authors of Graph Neural PDE Solvers with Conservation and Similarity-Equivariance (Masanobu Horie et al.) utilized dimensionless ratios of velocity to grid size to allow graph neural networks to generalize to spatial scales and orientations far beyond the training set.
- Surrogate Robustness: By removing unit-specific noise and arbitrary magnitudes, dimensionless equations allow surrogate models to focus entirely on the geometric relationships of the system. This has been demonstrated across diverse engineering domains, including helicopter transmission shafts (Vasiliki C. Panagiotopoulou et al. in Physics-informed surrogate modeling for a damaged rotating shaft), automotive aerodynamics (Ye Liu et al. in DragSolver: A Multi-Scale Transformer for Real-World Automotive Drag Coefficient Estimation), and core transport solvers for nuclear fusion (P. Rodriguez-Fernandez et al. in Enhancing predictive capabilities in fusion burning plasmas through surrogate-based optimization in core transport solvers).
Test-Time Projection and Policy Stability
When models must operate in uncalibrated environments, dimensionless invariants offer a powerful mechanism for zero-shot adaptation. The authors of Buckingham $\pi$-Invariant Test‑Time Projection for Robust PDE Surrogate Modeling (Seokki Lee et al.) demonstrated that out-of-distribution inputs can be projected back onto the training distribution's scale using physical invariants, cutting prediction errors by over 90% without retraining.
In reinforcement learning (RL) controlling physical systems, this representational strategy prevents hyperparameter sensitivity. As shown by Jeffrey F. Monaco et al. in Automated Aircraft Recovery via Reinforcement Learning: Initial Experiments, expressing state variables as dimensionless quantities ensures that the learned policy remains stable and strictly compliant with physical constraints across varying flight regimes and aircraft scales.
Dimension Perception in Large Language Models
Remarkably, the benefits of dimensionless representations extend beyond numerical simulators into discrete quantitative reasoning within large language models (LLMs). LLMs frequently fail math and physics problems when numbers are bound to physical units because they over-rely on specific unit-value pairs memorized during pre-training.
To resolve this, Yuncheng Huang et al. in Enhancing Quantitative Reasoning Skills of Large Language Models through Dimension Perception developed a framework that augments training data by swapping units within the same dimension (e.g., replacing meters with feet) while scaling numerical values accordingly. This forces the LLM to decouple raw numbers from specific units and perceive the underlying dimensional relationships.
Furthermore, when AI agents are tasked with scientific discovery, an inductive bias toward dimensionless ratios yields superior results. In Can Language Models Discover Scaling Laws?, Haowei Lin et al. showed that prioritizing formulas built from dimensionless ratios allows automated agents to discover scaling laws that are more robust and physically consistent than those previously derived by human experts.
Key Foundational Work
Real-Time Control of a Tokamak Plasma Using Neural Networks (1994) — contributed 2 concept elements to this direction.
Automated Aircraft Recovery via Reinforcement Learning: Initial Experiments (1997) — contributed 1 concept elements to this direction.