Dimensional analysis is the oldest feature-engineering trick in physics. If a quantity depends on $n$ measured variables built from $m$ base units (meters, seconds, kilograms), the Buckingham $\Pi$ theorem says it can be rewritten as a function of only $n-m$ unitless ratios, such as the Reynolds number. Models trained on these ratios transfer across scales and unit systems where raw-input models break. For example, a gradient-boosting model trained on 1-inch laboratory pipes reached $R^2=0.99$ on 12-inch Prudhoe Bay field data, while a mechanistic model reached 0.47 (Kanin et al.). Two kinds of redundancy survive this reduction, however. The theorem fixes how many ratios there are, but not which ones: any invertible recombination of a valid set is equally valid. And in practice the target usually depends on only $k<n-m$ combinations. The rest are dead weight, or are confounded with each other in the training data.
We propose a closed-form method that removes both redundancies at once. The method takes logarithms of the variables, restricts attention to the exponent directions that are exactly unitless, and fits any smooth predictor in those coordinates. It then ranks directions by how much the prediction actually changes along them, rather than by how much the data varies along them. The top eigenvectors of that sensitivity matrix are the governing dimensionless groups. A determinantal column-selection step then reads off the smallest set of raw sensors needed to form those groups. Comparing the sensitivity matrix with the data covariance also produces a diagnostic for when two groups cannot be told apart from the available experiments. That diagnostic is the case that silently breaks lab-to-field transfer.
Core Claim
Our claim is that a relevance-ranked spectral analysis restricted to the null space of the units matrix recovers the governing dimensionless groups as reliably as non-convex neural or symbolic searches, without restarts and at a small fraction of the compute. It also gives two outputs those searches do not: a minimal dimensionally closed sensor subset, and an identifiability warning for confounded groups.
Setup
Let $D\in\mathbb{R}^{m\times n}$ be the units matrix, where entry $(i,j)$ is the power of base unit $i$ in variable $j$. Let $B\in\mathbb{R}^{n\times p}$ be an orthonormal basis of its null space, with $p=n-m$. For strictly positive variables $x$, the coordinates $z=B^\top \log x$ are, up to a fixed rotation, the logarithms of a complete set of $\Pi$ groups. Every direction $w\in\mathbb{R}^p$ defines the dimensionless group $\exp(w^\top z)$, whose exponent vector on the raw variables is $Bw$.
We fit any smooth regressor $\hat g(z)$ to the dimensionless target, for example kernel ridge regression or a small neural network. We then form the expected gradient outer product:
$$C=\mathbb{E}\left[\nabla_z \hat g(z)\,\nabla_z \hat g(z)^\top\right],$$
where the expectation is taken over training points. The top-$k$ eigenvectors $W$ of $C$ span the directions the target responds to. The matrix $BW$ gives their exponents on raw variables. Dimensional homogeneity holds exactly by construction, because $D\,BW=0$. Because $B$ is orthonormal, the recovered subspace does not depend on which textbook basis one starts from. This resolves the basis non-uniqueness without a search. A final rounding step maps $BW$ to small-integer or rational exponents inside the null space and keeps them only if prediction error does not degrade.
The identifiability diagnostic compares $C$ with the data covariance $\Sigma_z$ of $z$. Suppose a relevant direction lies where $\Sigma_z$ has near-zero variance, or two relevant directions are nearly collinear under $\Sigma_z$. In a lab, for instance, Reynolds and Froude numbers often move together because both scale with velocity. Then no method can separate them from these data, and the method reports the confounded pair instead of committing to one.
For sensor reduction, the rows of $BW$ assign each raw variable a leverage score. Following Belhadji, Bardenet and Chainais, who prove that a projection determinantal point process (a sampler that prefers diverse, non-collinear subsets) with kernel $VV^\top$ selects columns with error bounds governed by effective sparsity, we sample raw-variable subsets from the projection kernel built on $BW$. We accept a subset only if the recovered exponent subspace still lies in the span of the units it contains. This "dimensional closure" constraint is exactly what statistical pruning lacks.
Relation to Prior Work
Two existing methods attack the non-uniqueness of $\Pi$ bases directly. Dimensionally Consistent Learning with Buckingham Pi learns exponents with a log–linear–exp network (BuckiNet), constrained optimization, and dimensionless sparse regression. Its authors report a non-convex landscape that needs restarts, and they add no mechanism to keep multiple learned groups from collapsing onto each other. Hi-π stacks dimensional analysis, PySR symbolic regression and polynomial mapping. It rediscovers Ra and Pr where PyDimension fails, and reaches 70% accuracy on the Karman–Tsien correction, but at high search cost and with large data needs. Both papers supply our evaluation tasks and serve as our main competitors. Our method replaces their search with one regression fit and one eigendecomposition.
Hi-π can find non-power-law combinations that a linear subspace in log space cannot represent. We treat that as a known limitation and test it on Karman–Tsien rather than hide it.
The redundancy literature motivates the relevance ranking. Gula et al. show that leading variance components of deep features capture background structure rather than the signal of interest. This is why we rank by gradient sensitivity and not by $\Sigma_z$. Zou and Adams use determinantal priors to stop latent components from redundantly encoding the same pattern, and Belhadji et al. supply the column-selection sampler we reuse. Neither applies a units constraint. Supervised pruning such as sparse mRMR can drop viscosity because it is weakly correlated with the target alone, which makes the Reynolds number impossible to form. This is the failure our closure constraint targets.
Field-level methods are a different regime, and we stay out of it. DimOL builds unit awareness into neural operators. Buckingham π test-time projection (Buckingham $\\pi$-Invariant Test‑Time Projection for Robust PDE Surrogate Modeling) maps out-of-range inputs onto training equivalence classes, using hand-derived groups and spatial means to avoid zeros. Our method could supply the groups the projection paper takes as given, but we restrict scope to scalar global parameters.
The log transform requires positive inputs, the same constraint BuckiNet has. Signed variables enter only through positive characteristic magnitudes plus a sign feature. We state this restriction rather than claim to solve it.
Research Program
The work has three parts. First, we build the estimator end to end: null-space coordinates, a smooth fit with cross-validated regularization, the gradient outer product, rational rounding, the identifiability diagnostic, and closure-constrained determinantal sensor selection. Second, we analyze when the estimator is consistent. The intuition is that if $\hat g$ converges in gradient, $C$ converges and its eigen-gap controls subspace error. We also characterize how the confounding diagnostic relates to the smallest eigenvalue of $W^\top\Sigma_z W$. Third, we test on tasks with known answers, then on two real industrial datasets where the questions are sample efficiency, transfer and sensor count rather than in-distribution accuracy. The in-distribution caveat matters: Jarmatz et al. report $Q^2=0.88\pm0.09$ with seven textbook dimensionless numbers versus $0.90\pm0.08$ with raw dimensional inputs. Dimensionless features did not buy accuracy there, so we do not promise it.
Baselines
The naive reference is a random forest or gradient-boosting model on raw dimensional inputs with standard scaling. On the fouling data it reaches the $Q^2=0.90\pm0.08$ reported by Jarmatz et al.. It represents doing nothing. We expect it to match us in distribution and to lose under range shift and unit rescaling.
The textbook-$\Pi$ baseline feeds hand-chosen groups to the same learner: the seven numbers of Jarmatz et al. ($Q^2=0.88\pm0.09$), and the Reynolds and Froude features of Kanin et al. ($R^2=0.95$ in the lab, $0.99$ on Prudhoe Bay). We beat it if we match its accuracy using fewer groups and fewer raw sensors, or if we recover its groups without being told them.
The strongest method for discovering groups is BuckiNet from Dimensionally Consistent Learning with Buckingham Pi. It optimizes the same objective we approximate in closed form. No reported quantitative result was found on Lacuna for BuckiNet; that paper reports only qualitative recovery on the rotating hoop, Blasius and Rayleigh–Bénard tasks. We give it the same training data and a wall-clock budget equal to ten times ours, spent on random restarts. We count our method ahead if it reaches equal or lower subspace error with no restarts, or lower error at equal compute.
For Hi-π, the reported figure is 70% discovery accuracy on Karman–Tsien. We use its own protocol there and on its Rayleigh–Bénard and Nikuradse tasks. We expect to lose on Karman–Tsien, because that relationship is not a power law. Matching Hi-π on Rayleigh–Bénard and Nikuradse at much lower cost would count as a win.
Three further baselines target redundancy specifically. The first is PCA on the null-space coordinates $z$, which isolates relevance-versus-variance ranking. The second is mRMR over raw variables and over candidate groups. The third is the same gradient-outer-product method applied to unconstrained $\log x$, which isolates the contribution of the null-space restriction. For sensor selection we compare determinantal sampling against greedy top-leverage selection and against mRMR, all at matched subset size.
Why not just…?
Why not just run PCA or an autoencoder on the inputs? Neither respects units, so their components mix meters with seconds and cannot transfer across scale. Even PCA restricted to the null space ranks by variance. Gula et al. show variance tracks background rather than signal. We expect the same here wherever the experiment designer varied an irrelevant group widely. Our null-space PCA ablation measures this directly on the Rayleigh–Bénard data, where the true groups are known.
Why not just feed all $n-m$ null-space groups to gradient boosting? This is a real threat for in-distribution prediction. Jarmatz et al.'s near-tie suggests it may win on that metric alone. It gives no minimal set of groups, no sensor reduction, and no warning when groups are confounded. Its extra irrelevant inputs should cost accuracy in small-sample regimes and under range shift. Learning curves at 50 to 500 points and a disjoint-range split will show whether that cost exists. If it does not, our contribution narrows to interpretability and sensor count, and we will report that.
Why not just run BuckiNet with more restarts? Given enough restarts, it may find the same subspace. The question is cost and reliability. The compute-matched comparison above settles it. The ablation without the null-space restriction separately shows whether the gain comes from the dimensional constraint or from the closed form.
Why not just prune raw features with mRMR or Lasso first? These methods can discard a weakly correlated variable that is needed to cancel units, such as density or viscosity in the Reynolds number. We will count how often mRMR subsets fail dimensional closure for the true groups. We will also compare its held-out accuracy at equal subset size against closure-constrained determinantal selection.
Proposed Experiments
On the ground-truth tasks, we measure the principal angle between the recovered and true exponent subspaces, the rate of exact rational exponent recovery, the sample size needed for 90% recovery, and wall-clock time. The rotating hoop, Blasius boundary layer and Rayleigh–Bénard tasks follow BuckiNet's setup. Nikuradse and Karman–Tsien follow Hi-π's, with added noise levels. On Karman–Tsien we expect partial failure. Reporting where a log-linear subspace stops being adequate is part of the result.
A synthetic confounding test plants two relevant groups that are collinear in the training design and separated at test time. The diagnostic must flag them before deployment, and the test error after the shift should line up with the flag.
On the fouling dataset, we report $Q^2$ against the 0.90 and 0.88 references. We also report the number of groups and raw sensors retained, and accuracy on held-out temperature and flow-rate ranges.
On the multiphase pipe data, if the lab and field records are obtainable, we train on lab data only. We test zero-shot on the field data against the reported $R^2$ of 0.99 on Prudhoe Bay, and ask whether our recovered groups coincide with Reynolds and Froude. If the field data are unavailable, we instead split the lab set by pipe diameter.
The decisive checks are three. Does relevance-ranked null-space spectral analysis beat null-space PCA at recovering known groups? Does it match restart-heavy BuckiNet at a tenth of the compute? Does closure-constrained sensor selection keep accuracy at subset sizes where mRMR breaks dimensional closure? Failing the first two would mean the closed form adds nothing over existing search methods.
References
- A predictive model for steady-state multiphase pipe flow: Machine learning on lab data
- A determinantal point process for column subset selection
- Dimensionally consistent learning with Buckingham Pi
- Hierarchical Dimensionless Learning (Hi-π): A physics-data hybrid-driven approach for discovering dimensionless parameter combinations
- Gaussian Image Anomaly Detection with Greedy Eigencomponent Selection
- Priors for Diversity in Generative Latent Variable Models
- Sparse minimum Redundancy Maximum Relevance for feature selection
- DimOL: Dimensional Awareness as a New 'Dimension' in Operator Learning
- Buckingham $\pi$-Invariant Test‑Time Projection for Robust PDE Surrogate Modeling
- Development of a soft sensor for fouling prediction in pipe fittings using the example of particulate deposition from suspension flow