The split Hopkinson pressure bar is the standard way to measure how materials behave when struck hard and fast. A projectile hits a long elastic bar, a stress wave travels into a small specimen, and strain gauges on the bars record three waves: the incident wave, the part reflected back from the specimen, and the part transmitted through it. The specimen's stress-strain curve is then computed from these three signals using one-dimensional wave theory. That theory assumes the specimen is loaded evenly and deforms without friction against the bar ends. Finite-element studies show both assumptions fail. Friction and radial inertia inflate the measured flow stress by 9–19%, and specimen shape or slenderness can shift it by up to 30%. As a result, part of what the field reports as "strain-rate strengthening" is really an artifact of the test setup.
We propose a neural network that takes the three raw bar waves and outputs a corrected stress-strain curve. It is trained mostly on a large library of finite-element simulations. The key ingredient is a second training target that is observable in both simulation and the real laboratory: the time history of the specimen's radial bulging, which high-speed cameras routinely record. We use this bulging prediction as an auxiliary task, meaning an extra prediction head used only during training and discarded at test time, as auxiliary-task methods in vision do. Because bulging shape directly reflects friction, it steers the shared encoder to represent the nuisance mechanics rather than absorb them into the stress estimate. Because it is measurable on real specimens, it also gives a supervised path from simulation to real tests, where the true stress is never available.
Core Claim
Auxiliary regularization in vision usually rests on the argument that labels are scarce. That argument is weak here, since a simulator can generate as many labeled pulses as we want. Our claim is different. An auxiliary target that is physically tied to the corrupting mechanism (friction-driven barreling) and observable in the deployment domain improves robustness under shift, not in-distribution accuracy. We predict gains on held-out specimen geometries, unseen constitutive families, and real tests. We also predict little or no gain on held-out simulations drawn from the training distribution. If the auxiliary head helps only in-distribution, or not at all, the claim is falsified.
Relation to Prior Work
The motivating evidence comes from Numerical and Theoretical Analysis of the Inertia Effects and Interfacial Friction in SHPB Test Systems. That work uses ABAQUS/Explicit simulations to quantify the friction, shape and slenderness errors, and derives optimal specimen aspect ratios. It stops at forward simulation and design advice; it builds no inverse model that corrects a given measurement. We reuse its simulation setup, covering cylinders and cubes, slenderness from 0.5 to 1.5, friction coefficients from 0 to 0.5, impact velocities of 3 and 10 m/s, and Johnson-Cook and Zhu-Wang-Tang materials, as our data generator.
The method comes from three auxiliary-task papers.
- Retina U-Net adds a full-resolution segmentation loss to a detector and finds the largest benefit with few training cases.
- NurtureNet regularizes scalar body-measurement regression with pseudo-labeled masks and keypoints. - Dense Road Surface Grip Map Prediction regularizes friction prediction with auxiliary water, ice and snow thickness heads, the closest analogue to a physically meaningful auxiliary state.
All three operate on images of static or slowly changing scenes, and all three use auxiliary targets available only at training time. Ours is a transient wave inversion, and the auxiliary target is observed at deployment, where it provides labels in the real domain that the primary task lacks.
The constitutive-modelling papers constrain what the output should be. A Constitutive Relation Based on the Johnson–Cook Model for Ti-22Al-23Nb-2(Mo, Zr) and Strain-Rate Effect on Anisotropic Deformation Characterization of High-Strength Aluminum Alloy Sheet show that fitted constitutive equations need hand-built extra terms. If the input curves carry setup artifacts, those terms may be fitting the apparatus rather than the material. We therefore output the corrected curve itself rather than model parameters, so that any constitutive model can be fit afterwards.
Research Program
We first build a simulation library of tens of thousands of runs. Each run samples geometry, friction, striker velocity and a material law. The material laws include Johnson-Cook, Zhu-Wang-Tang viscoelastic, and a softening variant in the spirit of the Ti2AlNb correction. From each run we record:
- the three bar-gauge waves;
- the true volume-averaged specimen stress and strain;
- the radial surface displacement profile along the specimen height over time, sampled as a camera would see it.
The model has a one-dimensional convolutional or recurrent encoder over the three waves, plus a scalar input for nominal geometry. A primary head outputs the corrected stress-strain curve. An auxiliary head outputs the bulging profile. Training minimizes
$$\mathcal{L} = \mathbb{1}_{\text{sim}}\,\lVert \hat{\sigma} - \sigma \rVert^2 + \lambda\,\lVert \hat{r} - r \rVert^2,$$
where $\hat{\sigma}$ and $\sigma$ are the predicted and true stress-strain curves, $\hat{r}$ and $r$ the predicted and observed radial profiles, $\lambda$ a weight tuned on validation, and $\mathbb{1}_{\text{sim}}$ equal to one for simulated samples and zero for real tests. Real tests therefore contribute only through the bulging term, which updates the shared encoder.
Using real data to update the encoder only through the bulging loss could let the stress output drift, a form of negative transfer (the auxiliary task pulling the shared encoder in a direction that hurts the main task). We therefore freeze the primary head during real-data adaptation, and we check that performance on simulated data does not degrade.
We also run a real campaign. One aluminum alloy and one mortar are each tested at several slenderness ratios, with and without end lubrication, and with both cylindrical and cubic specimens. Every test is filmed by a high-speed camera. This setup assumes camera access, which is a real constraint on adoption.
Baselines
Naive reference. Standard one-dimensional three-wave processing. Its error on simulated data is what the corrected methods must beat. The inertia-and-friction study reports 13% (3 m/s) and 9% (10 m/s) stress inflation for rate-independent materials as friction goes from 0 to 0.5, 19% and 16% for viscoelastic materials, and up to 30% for cubes versus cylinders. We will reproduce these on our library as a sanity check before comparing anything else.
Classical correction. Analytical Davies–Hunter-type inertia and friction formulas applied to the same pulses. They need an assumed friction coefficient, so we give them the true simulated value. This is a generous handicap in their favour.
Strongest competitor: single-task network. An identical encoder and primary head, trained on the same library with the same compute and hyperparameter-search budget, but without the bulging head. It is strong because simulation removes data scarcity, so auxiliary regularization might add nothing.
Target ablation. The same auxiliary setup, but with the auxiliary head regressing the scalar friction coefficient instead of the bulging profile. This separates "any friction-related auxiliary signal" from "a dense, lab-observable one". It is also the only variant besides ours that could be adapted on real data, and even then only if friction were independently measured, which it usually is not.
Reference ceiling. Iterative inverse finite-element calibration, which repeatedly re-runs ABAQUS, fitting material parameters and friction until simulated waves match measured ones. It is accurate but costs hours per test.
No reported result was found on Lacuna for any neural inversion of Hopkinson-bar pulses. Beating the single-task network therefore means lower stress-curve error on the out-of-distribution splits, with non-overlapping seed intervals. On real data it means lower cross-configuration spread, defined below.
Why not just…?
Why not just use the analytical friction and inertia corrections? They assume simple geometry and a known, constant friction coefficient. The simulation study shows the viscoelastic error changes with velocity and the cube error comes from corner wave convergence, which one-dimensional formulas do not describe. Our cube and viscoelastic splits test this directly, even with the formulas given the true friction coefficient.
Why not just train a single-task network on abundant simulations? This is the real threat. In-distribution we expect it to match our model, and we say so. The disagreement is about shift. On an unseen material law or on real tests, the single-task network has no signal telling it which part of the waves is friction, and no way to use real data at all. If the single-task network matches ours on the held-out-family split and on real-data spread, the auxiliary head is unnecessary and we will report that.
Why not just run inverse finite-element calibration for every test? It is the accuracy ceiling. However, it is slow and requires choosing a constitutive form in advance, which the Ti2AlNb and aluminum papers show is itself error-prone. We report how close the network gets to this ceiling and at what speed-up, rather than claiming to beat it.
Why not just design better tests, using optimal slenderness and lubrication? Good practice reduces but does not remove the error, and it cannot fix tests already run. Our lubricated, optimal-slenderness configurations measure how much residual error remains.
Proposed Experiments
Simulated evaluation uses four splits:
- held-out runs from the training distribution;
- held-out friction range (trained on 0 to 0.3, tested on 0.3 to 0.5);
- held-out geometry (trained on cylinders only, tested on cubes);
- held-out material family (trained on Johnson-Cook, tested on Zhu-Wang-Tang).
The metric is error in the corrected stress curve, reported as root-mean-square error and average absolute relative error, the latter matching how the Ti2AlNb paper reports constitutive fit. The first decisive check is whether the gap between our model and the single-task network grows from the in-distribution split to the shifted splits. Our claim predicts it does. A second check trains on subsets of the library, in the spirit of Retina U-Net's small-data analysis, to separate robustness gains from data-efficiency gains.
On real data there is no true stress. We therefore use an invariance check. A material's true curve at a given strain rate should not depend on specimen slenderness, shape or lubrication. Raw processing should spread across these configurations by the tens of percent the simulations predict. A correct inverter should collapse that spread, and adapting with the bulging loss on real data should collapse it further than simulation-only training. As a guard against the model collapsing every curve to a single shape, we also check that the corrected curves still separate the two materials and still separate distinct strain rates.
The final check asks whether apparent strain-rate strengthening shrinks after correction. That result would directly quantify how much of the reported strengthening in the aluminum and concrete literature is an artifact of the test setup.
References
- Numerical and Theoretical Analysis of the Inertia Effects and Interfacial Friction in SHPB Test Systems
- Retina U-Net: Embarrassingly Simple Exploitation of Segmentation Supervision for Medical Object Detection
- NurtureNet: A Multi-task Video-based Approach for Newborn Anthropometry
- Dense Road Surface Grip Map Prediction from Multimodal Image Data
- A Constitutive Relation Based on the Johnson–Cook Model for Ti-22Al-23Nb-2(Mo, Zr) Alloy at Elevated Temperature
- Strain-Rate Effect on Anisotropic Deformation Characterization and Material Modeling of High-Strength Aluminum Alloy Sheet