The research, conducted by Anshul Agrawal, Yash Yalamalli, Siddhartha Kar, Vibha Sastry Hebbani Krishna, and Priyadarshan Venkatesh at M S Ramaiah Institute of Technology in Karnataka, India, focuses on micro electro discharge machining, or µEDM. Unlike conventional drilling, µEDM does not cut material with a rotating blade. Instead, it uses a sequence of tiny electrical sparks discharged across a narrow gap between an electrode tool and the workpiece. Each spark melts and vaporizes a microscopic volume of titanium, gradually eroding a hole. Because the process relies on thermal erosion rather than mechanical force, it can machine hard, tough alloys that would defeat ordinary drill bits. The trade-off is complexity: the final hole geometry depends on a web of interacting electrical and kinematic parameters that are difficult to model with classical physics alone.

The team systematically varied four key parameters during drilling experiments on titanium: capacitance, voltage, tool rotation speed, and feed rate. Capacitance and voltage together determine the energy stored and released in each discharge, while tool rotation speed and feed rate govern how the electrode advances and flushes debris from the gap. The outputs of interest were the entry diameter and exit diameter of each micro-hole—two quantities that together describe the hole’s size, taper, and overall geometric quality. Correlation heat maps and scatter plots revealed a strong linear dependence of both diameters on capacitance and voltage, along with significant interaction effects between them, while the rotational and feed parameters played comparatively smaller roles.

With the experimental dataset in hand, the researchers benchmarked an unusually broad lineup of models. Classical baselines included linear and quadratic regression, alongside support vector machines. The machine learning contenders spanned gradient boosting frameworks—LightGBM and XGBoost—neural networks built with multilayer perceptrons and TensorFlow, and support vector regression. The goal was to find which approach could most faithfully map the four process parameters to the two diameter responses across the full experimental space.

The verdict was clear. XGBoost, the extreme gradient boosting algorithm that builds ensembles of decision trees by iteratively correcting the errors of previous trees, delivered the highest predictive performance. On held-out test data it achieved coefficients of determination of 0.9884 for the entry diameter and 0.9421 for the exit diameter, meaning it explained nearly all of the variance in both responses. It also posted the lowest error metrics across the full dataset. Crucially, the authors did not rely on a single performance number to crown the winner. They applied the Friedman and Nemenyi statistical tests, which rank multiple models across repeated evaluations and determine whether differences in performance are significant rather than artifacts of chance, and they used Cohen’s d to quantify the effect size of the gaps between models. Together, these analyses established XGBoost’s statistical superiority over its rivals.

Accuracy alone, however, is not enough for engineers who must trust a model on a factory floor. A black box that predicts hole diameters without explaining why offers little guidance for process planning. To open the box, the team deployed explainable artificial intelligence techniques, including permutation feature importance and SHAP—SHapley Additive exPlanations—a method borrowed from cooperative game theory that assigns each input feature a fair share of credit for every individual prediction. The explainability analysis confirmed that capacitance and voltage were the dominant parameters controlling both entry and exit diameters, and it revealed that their influence was not simply additive: the two parameters interacted in nonlinear ways, so the effect of raising voltage depends on the capacitance level in play. Residual analyses, in which the differences between predicted and measured values are scrutinized for hidden structure, further validated the robustness of the model.

The physical story behind the statistics was corroborated by optical microscopy of the drilled holes. As capacitance and voltage increased, the images showed clear diameter enlargement, the formation of a recast layer—a thin rim of melted and resolidified titanium along the hole wall—and irregularities at the hole edges. These observations make intuitive sense: larger discharge energies carve wider craters and leave more molten material behind, which then freezes onto the machined surface. The microscopy thus connected the abstract feature-importance rankings to visible, measurable damage mechanisms, giving the machine learning model a firm physical footing.

Perhaps the most forward-looking element of the work is what the authors did after prediction. Building on the statistically validated XGBoost model, they constructed a surrogate-based inverse design system. In forward modeling, one supplies process parameters and predicts outcomes. Inverse design reverses the arrow: the engineer specifies the desired entry and exit diameters, and the system searches for the optimal combination of capacitance, voltage, tool rotation speed, and feed rate that will produce them. Because the surrogate model is fast to evaluate, optimization algorithms can probe thousands of candidate parameter combinations in seconds, a task that would require weeks of physical experimentation. The result is a practical workflow that moves manufacturing from trial-and-error toward specification-driven process planning: tell the system what hole you want, and it tells you how to machine it.

The significance of the study extends beyond titanium micro-drilling. Machine learning is increasingly woven into precision manufacturing, from in-situ monitoring of discharge states using radio-frequency signals to deep temporal networks that watch machining processes in real time. What distinguishes this work is its completeness: it pairs a competitive benchmark of multiple algorithms with formal statistical validation, layering of interpretability tools, physical verification through microscopy, and a functional inverse design capability on top. It also reflects a broader lesson emerging in the applied machine learning literature—that gradient-boosted tree ensembles often outperform deep neural networks on modest-sized tabular datasets of the kind typical in manufacturing experiments.

For industries that depend on micro-scale features in difficult materials—fuel injection nozzles, medical implants, aerospace cooling channels, and microfluidic devices among them—the framework offers a template for intelligent process planning. The authors have also made their datasets and analysis openly available through a public GitHub repository, lowering the barrier for other groups to reproduce, extend, or adapt the approach to new materials and machines. As manufacturing continues its shift toward data-driven decision-making, studies like this one suggest a future in which the spark that erases a few cubic micrometers of titanium is guided not by an operator’s intuition, but by a model that has learned, statistically and transparently, exactly how each parameter shapes the hole it leaves behind.

Subject of Research: Machine learning prediction and inverse design of micro-hole diameters in µEDM drilling of titanium

Article Title: A machine learning framework for prediction and inverse design of micro-hole diameters in titanium drilled by µEDM

Article References: Agrawal, A., Yalamalli, Y., Kar, S., Krishna, V. S. H., & Venkatesh, P. (2026). A machine learning framework for prediction and inverse design of micro-hole diameters in titanium drilled by µEDM. Neural Computing and Applications, 38(19), Article 801. https://doi.org/10.1007/s00521-026-12494-1

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12494-1

Keywords: machine learning, µEDM, titanium, micro-hole drilling, XGBoost, inverse design, explainable AI, SHAP, manufacturing, regression, process optimization, recast layer