The researchers’ target was a look-up table, the workhorse metamodel of aerodynamic simulation. A look-up table stores values of an output, such as a force or moment coefficient, at a lattice of grid points spanning several input variables, and interpolates linearly between them at runtime. Flight simulators and guidance algorithms love look-up tables because they are fast, transparent, and easy to evaluate millions of times per second. But building one from real flight data is deceptively hard. The team’s dataset contained more than 200,000 records, yet those records covered only part of the two-dimensional variable space and were distributed wildly unevenly across the grid cells. Some cells were stuffed with data; others were empty. That imbalance is poison for the classical least squares method, the 230-year-old linear technique that underpins most system identification, because it makes the optimization problem ill-conditioned or even unsolvable.

The mathematics of the problem lives in the Hessian matrix, the matrix of second derivatives of the loss function that least squares must invert to produce its one-shot analytical solution. When data is spread evenly, the eigenvalues of this matrix cluster together and the inversion is stable. When data clusters unevenly, the eigenvalues spread apart, and the ratio of the largest to the smallest can explode. In the study’s demonstration, the largest eigenvalue reached about 4,000, and without regularization the ratio could climb to values that made coefficient estimates wildly inaccurate. Worse, cells with no data at all produced rows of zeros in the Hessian, rendering the matrix singular and the inversion impossible. The coefficients that emerged in such cases were physically absurd: peak values roughly four times the highest value in the raw data, sitting at grid points adjacent to empty cells, artifacts of the model faithfully chasing noise in sparse regions.

The team’s remedy for the coefficient stage was ridge regression, the simplest of the non-smoothness penalty techniques. By adding a penalty term proportional to the sum of squared coefficients, weighted by a constant alpha, ridge regression shrinks all parameters toward zero and guarantees that the Hessian remains invertible even where data is absent. Eigenvalues far below alpha are lifted to approximately alpha, taming the condition number, while eigenvalues far above alpha are barely affected. The trick is choosing alpha. Too small, and the model overfits, exhibiting low bias error but high variance error, meaning it drifts far from the true optimum by chasing noise. Too large, and the model underfits, smoothing away genuine physics. The researchers swept alpha across 21 powers of two, from 2 to the minus 10 up to 2 to the 10, and watched two competing indices move in opposite directions.

The first index was familiar: R squared, the statistical measure of how well the metamodel matches the raw data, which quantifies bias error. The second was more novel: a smoothness index computed from the coefficients themselves. Because real-world aerodynamic phenomena rarely involve instantaneous jumps, a good discrepancy model should vary gently across the grid. The team defined smoothness two ways, once using first derivatives, or gradients, of the coefficients, and once using second derivatives, or curvatures, averaged over all grid points and variables after normalizing each variable by its range. As alpha increased, R squared fell while both smoothness measures improved, tracing out a Pareto front of the bias-variance trade-off. Notably, the curvature-based measure proved more discernible than the gradient-based one, declining in a pattern that made changes in variance error easier to read, and it ultimately became the team’s preferred variance indicator.

The sweep also produced a striking diagnostic: the ratio of the maximum and minimum optimized coefficients to the maximum and minimum of the raw data. At small alpha these ratios were enormous, confirming that the solution had wandered far from anything physical. Around alpha equal to 2 to the 4, or 16, the ratios approached zero, meaning the coefficients had shrunk to the same order of magnitude as the data itself. The team recommended this value as the optimum for the fixed-grid stage, achieving a smoothness index dramatically reduced while R squared remained near 68 percent. That figure may look modest for a metamodel, but the authors attribute the ceiling to high noise levels in the flight data and to a deliberately small 6-by-6 grid of 36 coefficients. Their goal was not to maximize R squared but to characterize how it trades against smoothness, and on that question the parameter study delivered a clear answer.

The second stage of the method is where the design becomes genuinely clever. Rather than accepting a fixed, equally spaced grid, the researchers let the grid points themselves move, optimizing their positions to reduce bias error further. Moving grid points makes the problem nonlinear, so they employed gradient-based constrained optimization in MATLAB, reformulating the problem so that the decision variables were the step sizes between adjacent grid points, with constraints keeping every step positive and the total span fixed. This nesting created a two-loop architecture: in the outer loop, a nonlinear optimizer adjusts grid positions and the ridge constant; in the inner loop, the linear least squares with ridge regression re-solves for all 36 coefficients at the new grid. The loss function for the outer loop combined one minus R squared with a curvature-based non-smoothness penalty weighted by a factor kappa, tested at values of 1.0, 0.5, 0.3, and 0.2.

The results of the grid optimization revealed both the power and the ironies of the approach. Across the weighting factors, grid optimization delivered substantial improvements in bias error, and in most cases improved smoothness as well, compared with the fixed-grid parameter study at comparable alpha. But the interplay between the two penalty constants produced counterintuitive outcomes. When kappa was set to 1.0, placing maximum emphasis on smoothness, the optimizer settled on a smaller ridge constant of about 16, and smoothness actually deteriorated slightly. Conversely, at kappa equal to 0.2, the least emphasis on smoothness, the optimizer pushed the ridge constant up sharply and achieved the largest reduction in smoothness. Meanwhile, the grid itself adapted meaningfully: as the weight on smoothness decreased, grid points migrated toward the boundary regions of the variable space, coefficient peaks grew there, and bias error fell at the cost of variance. The lesson is that the two regularization knobs do not act independently, and the cheap linear parameter study proved an effective way to choose kappa before committing to the expensive nonlinear stage.

What makes this work resonate beyond aerodynamics is its economy. The authors emphasize that the nested process of classical methods, least squares inside gradient-based optimization, consumed little computational resources and time, especially because the decisive parameter study relies only on the linear inner loop with a fixed grid. In an era when machine learning models with millions of parameters are routinely proposed for system identification, there is something almost radical about a 36-coefficient model, built with two centuries-old mathematics plus a well-placed penalty, that demonstrably balances bias and variance on real flight test data. The technique also preserves consistency with the original aerodynamic prediction model, which was itself built as a look-up table during the design phase, so the discrepancy layer can be merged directly into existing simulation pipelines.

The implications ripple outward. Accurate discrepancy models sharpen performance estimates, improve simulator fidelity, and feed corrected aerodynamics back into the design of future vehicles, a loop the authors expect to benefit future modeling and simulation products. More broadly, the study is a demonstration that the bias-variance decomposition, a cornerstone of statistical learning, can be operationalized with smoothness as a practical, computable proxy for variance error, and that curvature beats gradient as the smoother’s measuring stick. As flight test campaigns grow and vehicles fly further from their predicted envelopes, methods like this one, quiet, disciplined, and mathematically honest about what sparse data can and cannot support, may become as essential to aerospace engineering as the wind tunnels that started it all.

Subject of Research: Construction of multivariable look-up table aerodynamic discrepancy models using nested linear and nonlinear optimization with non-smoothness penalties

Article Title: Multivariable Look-up Table Construction for Aerodynamic Prediction Discrepancy Model Using Nested Grid Optimization of Linear and Nonlinear Methods with Non-smoothness Penalty

Article References: Jung, S. Y., Nam, H., & Yim, K. (2026). Multivariable Look-up Table Construction for Aerodynamic Prediction Discrepancy Model Using Nested Grid Optimization of Linear and Nonlinear Methods with Non-smoothness Penalty. International Journal of Aeronautical and Space Sciences. https://doi.org/10.1007/s42405-026-01299-0

Image Credits: AI Generated

DOI: 10.1007/s42405-026-01299-0

Keywords: aerodynamics, look-up table, system identification, ridge regression, bias-variance trade-off, flight test data, metamodeling, optimization, regularization, discrepancy model, least squares, aerospace engineering

The work, led by Nikita Grewal and Yash Arora together with S. K. Gupta and Sanjeev Kumar, builds on the twin support vector machine, a widely recognized classification model first introduced as a faster, leaner cousin of the classical support vector machine. Rather than searching for a single separating hyperplane, the twin version constructs two nonparallel hyperplanes, one for each class, and assigns each point to whichever hyperplane it sits closer to. This structural difference typically makes training faster than the classical formulation, since each optimization problem involves the samples of only one class at a time. But the twin model has well-documented weaknesses: it is sensitive to noise and outliers, and when class sizes are unequal it can produce systematically biased decision boundaries that favor the dominant class.

The Roorkee team’s central insight is that not all samples deserve equal trust. In real datasets, some points sit comfortably in the heart of their class while others linger near the boundary or belong to scattered secondary clusters, and a few are outright errors. To capture this heterogeneity, CPLTSVM assigns each training sample a fuzzy membership weight derived from its proximity to the nearest class centers. Instead of assuming every class forms a single compact blob, the model adopts a polycentric view: each class may have several centers, reflecting the multi-cluster structure that naturally arises in high-dimensional data. Samples close to any center of their own class receive high membership values, meaning they exert strong influence on the learned hyperplane, while distant or structurally isolated points receive lower weights and therefore contribute less to the final decision rule. This mechanism embeds global structural information about the dataset directly into the training objective, something conventional support vector formulations ignore entirely.

The fuzzy weighting scheme draws on a long intellectual lineage. Fuzzy sets, introduced by Lotfi Zadeh in 1965, allow data points to belong to categories to intermediate degrees rather than in a binary fashion, and researchers have spent decades folding such gradations into kernel-based classifiers. Earlier fuzzy twin support vector machines have used information entropy, affinity measures and intuitionistic fuzzy memberships to dampen the effect of noisy points. What distinguishes the new approach is the combination of two complementary signals. The polycentric membership captures how representative a point is of its class structure, while a separate class probability estimate captures how confidently a point belongs to its assigned class at all.

That probability estimate comes from the k-nearest-neighbor technique, one of the oldest and most intuitive tools in pattern recognition. For each training sample, the model examines its closest neighbors and computes the fraction that share its label, yielding a local estimate of class membership probability. Points surrounded by neighbors of the same class earn high probability scores; points in mixed neighborhoods earn low ones. When these probabilities are folded into the fuzzy membership function, noisy samples and borderline outliers are automatically down-weighted, reducing their power to drag the decision boundary away from the true underlying structure. The result is a classifier that is robust by construction rather than by post-hoc correction.

Computational efficiency was a second design priority. Standard twin support vector machines solve two quadratic programming problems, which can be expensive on large datasets. By integrating the fuzzy membership weights into the least squares twin support vector machine formulation, which replaces the inequality constraints of the original with equality constraints and a squared loss, the authors convert training into the solution of systems of linear equations. This dramatically lowers the computational cost while preserving the benefits of the fuzzy weighting scheme, making the approach practical for the kinds of large, messy datasets where imbalance problems are most acute.

To evaluate the framework, the team ran experiments on twenty benchmark imbalanced datasets, testing both linear and nonlinear kernel versions of CPLTSVM against several state-of-the-art competitors. Performance was measured with three metrics suited to imbalanced settings: the area under the receiver operating characteristic curve, known as AUC, the geometric mean of class sensitivities, known as G-mean, and the F-measure, which balances precision and recall. The numbers are striking. The model achieved an average AUC of 89.71 percent with the linear kernel and 94.87 percent with the nonlinear kernel, and it consistently surpassed the existing techniques across the benchmark suite. The authors supplemented raw scores with comprehensive statistical analyses to confirm that the observed gains were significant rather than artifacts of dataset selection.

The researchers also demonstrated the model’s practical value on a task with real clinical stakes: breast cancer diagnosis from histopathological images using the BreakHis dataset. On this medical imaging benchmark, CPLTSVM delivered superior performance compared with baseline methods, suggesting the technique could support real-world classification pipelines where false negatives carry severe consequences. Because histopathological datasets are naturally imbalanced and noisy, with subtle visual differences between benign and malignant tissue, this application exercises exactly the weaknesses the new membership scheme was designed to address. The authors have made the source code publicly available on GitHub, allowing other researchers to reproduce the results and extend the framework.

The significance of the work extends beyond a single benchmark victory. Machine learning systems are increasingly deployed in settings where rare events dominate the consequences: detecting defective components in manufacturing lines, flagging fraudulent transactions, spotting early signals of disease, and identifying anomalies in wind turbine gearboxes, a domain where twin support vector machines have previously found application. Every one of these settings shares the same structural pathology that the Roorkee team targeted, and a classifier that resists both noise and imbalance without requiring delicate resampling or costly retraining could simplify deployment considerably. Compared with popular alternatives such as the SMOTE synthetic oversampling technique or cost-sensitive learning frameworks, which modify the data or the loss rather than the classifier’s view of structure, the fuzzy membership approach works within the learning algorithm itself, adjusting how much each sample matters based on where it sits in the geometry of the data.

There are, as with any method, avenues for future work. The polycentric membership depends on identifying class centers, and the k-nearest-neighbor probability estimation introduces a hyperparameter in the choice of k, questions of sensitivity that practitioners will want to probe. The evaluation here focuses on binary classification, the native territory of twin support vector machines, and extending the framework to multicategory problems, as earlier fuzzy proximal approaches have done, would broaden its reach. Still, the combination of strong empirical results, computational efficiency and demonstrated medical utility marks CPLTSVM as a notable advance in the ongoing effort to make classifiers trustworthy when the world refuses to hand us balanced data. As algorithms increasingly inform decisions about health, security and safety, techniques that give quiet, underrepresented samples their due weight are not merely academic refinements; they are prerequisites for reliable artificial intelligence.

Subject of Research: A fuzzy least squares twin support vector machine using class probability and polycentric memberships to improve classification of imbalanced datasets

Article Title: Class probability and polycentric fuzzy least squares twin support vector machine for imbalanced data

Article References: Grewal, N., Arora, Y., Gupta, S. K., & Kumar, S. (2026). Class probability and polycentric fuzzy least squares twin support vector machine for imbalanced data. Knowledge and Information Systems, 68(1), Article 259. https://doi.org/10.1007/s10115-026-02862-7

Image Credits: AI Generated

DOI: 10.1007/s10115-026-02862-7

Keywords: imbalanced datasets, support vector machine, twin support vector machine, fuzzy membership, polycentric membership, class probability, least squares, machine learning, medical diagnosis, breast cancer, BreakHis dataset, classification