Physics-Guided Multi-Fidelity Surrogate Modeling for Natural Supercavity Geometry of an Axisymmetric Conical Cavitator

arXiv:2610.18886 · physics.comp-ph, cs.LG · Submitted 2026-10-06 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Physics-Guided Multi-Fidelity Surrogate Modeling for Natural Supercavity Geometry of an Axisymmetric Conical Cavitator".

Jane: The gist The proposed PG-EHK framework enables rapid, accurate, and physically consistent prediction of cavity profile across the full practical design space of (σ, αc,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: We’re looking at the title, "Physics-Guided Multi-Fidelity Surrogate Modeling for Natural Supercavity Geometry of an Axisymmetric Conical Cavitator". It tells you immediately that this isn't just some black-box machine learning model; it’s grounded in physics.

Jane: It suggests they’re using known physics to guide the AI, which is what makes these kinds of predictions much more reliable than models that are trained purely on data without any physical constraints.

Lu: The authors are Vinh Pham, Van Tu Nguyen, and Warn-Gyu Park from Pusan National University and Ho Chi Minh City University of Technology. Their background is clearly strong in both the mechanical engineering side—dealing with fluid dynamics—and the computational science side—building these complex surrogate models.

Meng: From a deployment standpoint, having someone from a university setting on this kind of complex fluid problem means they are trying to bridge the gap between theoretical physics and what actually works in high-speed vehicle design.

Lalam: The paper is about natural supercavitation behind conical cavitators, and the authors are using this framework to accurately predict three outputs: cavity length Lc, maximum radius Rc, and the local cavity profile r over position x.

Tom: The implication is that for designers dealing with these complex shapes—especially for torpedoes where you need to know the exact shape for body coverage assessment—this framework gives them a fast way to check designs without needing massive CFD runs.

The paper's summary: Jane: Okay, so they summarize the method by saying they take these three analytical models and treat them as low-fidelity trend functions, or LF functions, and then combine them with sparse high-fidelity CFD data using an EHK method.

Tom: That EHK method is the key here because it’s designed to handle datasets that aren't perfectly nested. It lets you take these low-level physical guesses and use the high-fidelity data to correct the errors where those guesses are off.

Lu: They build this hierarchical composite of EHK surrogates, which means they have three sequential layers. Layer one predicts drag coefficient Cd using just the cone angle alpha c and cavitation number sigma, using those three analytical formulas as inputs for a first surrogate.

Meng: And layer two takes that and estimates the cavity length Lc and maximum radius Rc based on alpha c, sigma, and then that predicted Cd from layer one. That’s how they keep things computationally light while still using the CFD data to guide it.

Jane: Then layer three builds a surrogate for the local cavity radius r over x, which depends on everything from alpha c and sigma down to Lc and Rc as well. It’s a very structured way of building complexity step by step.

Lalam: They state that this framework is required because no prior study had applied a multi-fidelity surrogate framework that simultaneously fuses multiple competing analytical supercavity models as non-level low-fidelity trend functions with sparse high-fidelity CFD data to predict natural supercavity geometry behind a cone cavitator with a variable angle.

Tom: So the main point is they're solving the problem of predicting geometry with variable angles by intelligently blending multiple physical theories, which is what makes the system robust across the whole design space.

The paper's improvements: Tom: They describe a few key improvements in their proposed PG-EHK framework. One major thing is that they’ve structured it as three distinct EHK layers to compute Cd, Lc, and Rc sequentially based on the inputs they have.

Jane: That structure itself seems like a big step because it manages the complexity of relating drag to geometry across different conditions—sigma and alpha c—in a very organized way.

Lu: The paper also focuses on an improvement where they incorporate Bayesian-estimated scaling factors into the framework. This means the AI system learns the optimal weighting for each of those three analytical models directly from the CFD training data, instead of relying on manual tuning by a human engineer.

Meng: That is huge for deployment because it automates what used to be a tedious process of figuring out which physical model was most trustworthy for a given set of inputs. If the data shows one formula is consistently better than another in a certain region, the AI learns that weighting automatically.

Lalam: They also suggest an adaptive infill sampling strategy for future work, where the system uses prediction variance to decide where to run new CFD evaluations next. The goal there is to get equivalent accuracy with fewer than fifteen CFD training points by focusing on areas of high uncertainty, like low sigma and large alpha c.

Tom: That makes sense from a machine learning efficiency viewpoint. Instead of just sampling randomly, the AI tells you exactly where you need more expensive CFD evaluations to get the most bang for your buck.

Conclusion: Jane: So to wrap up, this paper introduces the Physics-Guided Multi-Fidelity Surrogate Modeling for Natural Supercavity Geometry of an Axisymmetric Conical Cavitator, which uses a hierarchical composite EHK approach to predict cavity length, radius, and profile using just twenty high-fidelity CFD evaluations.

Tom: The results show that this PG-EHK surrogate achieves a sixty-one percent reduction in prediction error compared to the best individual analytical approach tested on five independent hold-out CFD cases. It’s pretty solid performance for a model built like this.

Lu: What this means practically is that you get accurate predictions for maximum cavity length and radius, and the local cavity profile shape as functions of cavitation number sigma and cone half-angle alpha c.

Meng: For someone shipping models to production, the implication is that you can replace slow CFD checks with this fast AI predictor, which is essential if you want to test many design variations quickly in an early stage workflow.

Lalam: The paper also notes a limitation—the authors admit they are building this framework by integrating three specific analytical models and sparse CFD data, so the accuracy is fundamentally tied to how well those initial analytical models represent the true physics across the whole design space.

Jane: That’s a fair point. Overall, it shows that physics-guided multi-fidelity architecture can be very effective for bridging theoretical models and high-fidelity numerical computations in areas like supercavity geometry prediction.

Tom: Yeah, this work on the PG-EHK framework is a strong piece of work because it moves beyond just fitting data to create a model that understands the underlying physics enough to make sense of those physical predictions.

Lu: It sets a precedent for how we can use these multi-fidelity methods to fuse competing analytical knowledge when no single theoretical solution exists, which is something I think is really important for complex fluid problems.

Meng: It’s a practical tool that helps reduce the computational cost of design iteration without sacrificing too much accuracy, and that's what matters when you're dealing with real-time applications.

Lalam: We’ll keep an eye on how they use the adaptive sampling strategy to make it even more efficient in future iterations.

Vinh Pham, Van Tu Nguyen, *Warn-Gyu Park

School of Mechanical Engineering, Pusan National University · Faculty of Applied Science, Ho Chi Minh City University of Technology (HCMUT)

physics.comp-ph, cs.LG

Submitted: 2026-10-06

Updated: 2026-10-06

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 84/100

The gist: The gist The proposed PG-EHK framework enables rapid, accurate, and physically consistent prediction of cavity profile across the full practical design space of (σ, αc, x) using only 20

Key concepts

PG-EHK Framework: This is a hierarchical surrogate modeling technique that combines multiple low-fidelity (analytical) models with high-fidelity (CFD) data. It uses Bayesian scaling factors to build a physical baseline, ensuring the resulting model is physically consistent across different design conditions.
Multi-Fidelity Modeling: This approach uses models of varying accuracy—low fidelity (simple analytical formulas) and high fidelity (complex CFD simulations). The PG-EHK framework intelligently fuses these different levels of detail to create a single, highly accurate prediction that is computationally cheap.
Non-Level Disagreement: This concept describes when different physical models, even if based on the same underlying physics, produce systematically different results. The study confirmed that three analytical models for cavity geometry showed this disagreement across the design space, justifying the need for a composite surrogate approach.
Cavitation Number ($\sigma$): The cavitation number is a key parameter representing the ratio of dynamic pressure to vapor pressure in fluid flow. It dictates whether cavitation occurs and significantly influences the resulting supercavity shape and drag coefficient.

Terminology

Fidelity — 1× in this paper · explained in 2 episodes

Summary

The gist The proposed PG-EHK framework enables rapid, accurate, and physically consistent prediction of cavity profile across the full practical design space of (σ, αc, x) using only 20 high-fidelity CFD evaluations

Problem Statement

The objective of this study is to construct a computationally efficient surrogate model that accurately maps the cavitator design variables of (σ, αc, x) to the primary geometric characteristics of a natural supercavity generated by conical cavitator in high-speed underwater flow The cavity geometry is parameterized by the cone angle betac and half-angle alphac = 0.5betac, with a fixed slant length, Lscl; Rn denotes the base radius of the cone cavitator The three primary outputs of interest are: (i) the normalized cavity length L̅c = Lc/Rn; (ii) the normalized maximum cavity radius R̅c = Rc/Rn, and (iii) the normalized cavity profile r̅= r(x)/Rn, which describes the radial extent of the cavity as a function of axial position x and is required for body coverage assessment in torpedo design The design space is spanned by three independent variables: (i) the cavitation number σ; (ii) the cone half-angle alphac; and the axial coordinate x measured along the symmetric axis of the cavity

Analytical Models and Disagreements

Three distinct analytical approaches for modeling natural supercavity geometry, those of Guzebsky, Choi and empirical Logvinovich approaches, are incorporated as LF baseline models of prior physical knowledge in the proposed PG-EHK framework All three are derived from the Logvinovich independence-section-expansion principle and the slender-body theory of axisymmetric cavity flows but differ in their closure assumptions and the treatment of the aft cavity region Their shared physics foundation makes them approximately correlated but systematically distinct in predictions, satisfying the non-level criterion for EHK application The comparison of analytical approaches against CFD reference data reveals that both Guzevsky and Choi formulas consistently show that Cd increases monotonically with σ for all cone angles, and that larger αc yield a higher drag coefficient at any given σ The empirical Logvinovich model severely and consistently overestimates both L̅c and R̅c across all cone angles and cavitation numbers

PG-EHK Framework Construction

The construction of the proposed PG-EHK model for the natural supercavity geometry behind a conical cavitator proceeds in four sequential steps Step 1 (Data acquisition) involves performing LF analytical analysis on a uniform grid of N = 400 points over the design space (σ × αc×ξ: 5 × 4 × 20) to generate LF surfaces for Cd, L̅c, R̅c and r̅(ξ) Step 2 (Inter-model disagreement analysis) compares the three traditional analytical approaches to the CFD data to confirm that they exhibit non-level disagreement across the design space, justifying the PG-EHK approach over single analytical approach Step 3 (PG-EHK surrogate construction) trains the proposed PG-EHK surrogate, incorporating LF analytical models and CFD data The framework is a hierarchical composite of EHK surrogates comprised of three sequential EHK layers that compute Cd, L̅̂c and R̅̂c, respectively, and r̅(ξ) as functions of the input variables (alphac, σ, ξ) Layer I constructs a surrogate of Cd with respect to the input of (alphac, σ) using three analytical drag coefficient formulas as LF models In Layer II, two surrogates are constructed to estimate L̅̂c and R̅̂c as functions of inputs (alphac, σ, Cd) using three analytical formulas for L̅c and R̅c In Layer III, a surrogate of the local cavity radius r̂̅ is constructed as a function of input (alphac, σ, L̅̂c, R̅̂c, ξ) using two Logvinovich-based formulas

Surrogate Validation and Accuracy

The PG-EHK surrogate achieves 61% reduction in prediction error compared to the best-performing individual analytical approach on cross-validation on five independent hold-out CFD cases In Figure 10, the PG-EHK surrogates consistently produce the closest agreement with the CFD reference profiles across all five held-out test cases, accurately capturing both maximum cavity radius and overall cavity profile shape For Cd, PG-EHK is approximately 37% more accurate than Kriging for the same quantities The PG-EHK surrogate achieves the lowest RRMSE and RMAE for all output quantities

Conclusion

The proposed PG-EHK framework demonstrates that the physics-guided multi-fidelity architecture of PG-EHK enables accurate and computationally efficient supercavity geometry prediction from a small CFD training dataset, making it well-suited for integration into early-stage design workflows <ref:489 These results demonstrate that the proposed PG-EHK enables accurate and computationally efficient super

Improvements for AI systems

textbfPhysics-Guided Multi-Fidelity Surrogate Framework for Geometry Prediction: Capabilities of PG-EHK system development and validation within a 20-case CFD training set across the full design space of (σ, αc, x). This improved AI system can predict maximum cavity length and maximum cavity radius and local cavity profile as functions of cavitation number σ and cone half-angle αc using fewer than 20 high-fidelity CFD evaluations while achieving a 61% reduction in prediction error compared to the best-performing individual analytical approach.
textbfAutomated Model Weight Learning for Analytical Formulas: Integration of Bayesian-estimated scaling factors into the PG-EHK framework. This enables the AI system to automatically learn optimal scaling weights for each analytical LF model from the CFD training data, as reflected in the estimated scaling factors reported in Table 2, ensuring that formulas known to carry systematic bias are automatically suppressing[ed] without requiring any manual tuning of model weights.
textbfAdaptive Infill Sampling Strategy: Implementation of variance-driven sampling for future training. The improved system can employ an adaptive infill sampling strategies driven by the PG-EHK prediction variance, with the objective of achieving equivalent or superior surrogate accuracy with fewer than 15 CFD training points by concentrating new evaluations in regions of high predictive uncertainty, particularly at low σ and large αc where analytical models are least reliable.
textbf Real-Time Cavity Shape Management: Development of an inverse PG-EHK formulation for closed-loop control. This capability allows the system to modulate the cone half-angle continuously to maintain a prescribed target cavity length Lc against disturbances in the ambient cavitation number σ, making it directly compatible with model-predictive control architectures operating at actuation timescales.

Sources

AAM-VDT: Vehicle Digital Twin for Tele-Operations in Advanced Air Mobility