De-Risking Robotic Welding for In-Space Metal Joining Through Physics-Grounded Digital Twins

arXiv:2610.21939 · cs.RO, physics.app-ph · Submitted 2026-10-15 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "De-Risking Robotic Welding for In-Space Metal Joining Through Physics-Grounded Digital Twins".

Dev: The gist The UK Space Agency-supported ISPARK project is developing a physics-grounded digital twin framework for structured process de-risking through robotic welding Framework Architecture The proposed architecture is organized into…

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

Paper summary: Rosa: So, this paper is about this thing called "De-Risking Robotic Welding for In-Space Metal Joining Through Physics-Grounded Digital Twins". It’s tackling the problem that robotic welding in space, especially in low gravity and low pressure, is still really not well understood because we don't have many flight tests yet.

Dev: Yeah, the core idea here is to build a physics-grounded digital twin framework. They’re using this twin to de-risk the process before you even try it on hardware. It claims they’ve developed a model-based prototype that combines several things: a finite-volume thermal weld-pool model, Gaussian processes, and neural network surrogates, along with a simulated six-axis manipulator.

Taro: So, what’s the big claim? They are using this digital twin to create some kind of structured process de-risking. That sounds like they want to screen the welding conditions before any actual experiment happens.

Rosa: Exactly. The paper evaluates this model-based prototype using a reduced thermal weld-pool model that includes solid-liquid phase change and effective thermocapillary transport, and a local-pressure parameter. They test how gravity affects things separately through a hydrostatic to capillary pressure ratio, but they explicitly state gravity does not enter the thermal energy equation itself.

Dev: And for the surrogates, they found that Gaussian processes have a lower normalized mean absolute error than neural networks for all eight outputs. Plus, in ten held-out boundary cases, seventy-four out of eighty solver responses fall within the nominal ninety percent GP predictive intervals.

Taro: That sounds like good fidelity for a surrogate model. So, when you look at the simulations, they also found that a three output Gaussian process query takes two hundred forty-nine microseconds to run, which is super fast when you consider the scaling and predictive variance involved.

Rosa: And they didn't stop there; in a simulated UR5 L-shaped-seam study, they used a proportional controller to adjust the synergic power source setting as the travel speed dropped near the corner. Finite-volume re-evaluation showed all twelve closed-loop operating points and nine of those twelve open-loop points fall within an illustrative one point zero to three point zero millimeter root side width band.

Dev: So, what does this mean for us? It supports using this framework to screen welding conditions and control strategies before you actually start experimental tests on hardware. But the authors are clear about what they didn't do; they say thermal-vacuum, targeted reduced-gravity tests, and hardware-in-the-loop studies are still needed to fully assess the model and the controller.

Taro: That makes sense for a physics model. They’re telling us this is a tool for screening, not a final answer. It highlights the gap between simulation and reality.

Rosa: Right, so they’ve built a physics-grounded digital twin that helps bridge that gap by using these advanced modeling and data science tools to screen welding conditions in space environments. This leads us into what this paper actually means for the world when we talk about in-space servicing.

Conclusion: Dev: So, wrapping up this discussion on "De-Risking Robotic Welding for In-Space Metal Joining Through Physics-Grounded Digital Twins", the authors are essentially presenting a framework that combines a reduced thermal weld-pool model, data-driven surrogates, and a robotic control simulation to screen candidate in-space arc welding conditions.

Rosa: The key finding is how they achieved this screening capability by using those physics models and then feeding the results into surrogate models that are fast and accurate enough for real-time use. They showed that the GP approach gave lower normalized mean absolute error compared to neural networks across all eight thermal and weld-pool outputs.

Taro: It means we can predict how a weld pool will behave under different conditions, like varying gravity or pressure, without needing to run those expensive, slow finite-volume calculations every single time.

Dev: That’s the practical implication for the engineering side. They showed that you can use this to screen welding conditions and control strategies before you even start experimental tests on hardware. It’s a way to save time and resources down the line by figuring out what works in simulation first.

Rosa: The whole thing is about taking a promising area, robotic welding in space, and using this digital twin approach to systematically reduce the risk associated with trying it out in actual flight environments where testing is hard. This work was supported by the UK Space Agency under National Space Innovation Programme two through the Intelligent SPace Arcwelding Robotic Kit project <ref:2610.21939#pg1,Intelligent SPace Arcwelding Robotic Kit project>.

Dev: So, for anyone listening who just wants a simple version, this paper is about using advanced modeling and data science to create a tool that helps us safely prepare robotic welding operations for in-space applications.

Zhou Daniel Haoa, Umer Masood Chaudrya, Robert Shaw, Andrew Cheney, Nick Ludford, Mark Simsb, Hongbiao Dong

University of Leicester Department of Science, School of Computing and Mathematical Sciences, TWI Ltd, University of Birmingham School of Metallurgy and Materials

cs.RO, physics.app-ph

Submitted: 2026-10-15

Updated: 2026-10-15

Comments: 14 pages, 7 figures. An alternative version was accepted for oral presentation at the 77th International Astronautical Congress (IAC 2026)

Code: https://github.com/UniversalRobots/Universal_

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

Importance score: 72/100

The gist: The gist The UK Space Agency-supported ISPARK project is developing a physics-grounded digital twin framework for structured process de-risking through robotic welding Framework Architecture The

Key concepts

Finite-Volume Thermal Weld-Pool Model: This is a physics simulation that models how heat moves through molten metal during welding. It tracks the energy transport, including melting and solidification (phase change), based on the arc heat input. Gravity is handled separately in this model.
Gaussian Process (GP) Surrogates: These are mathematical tools trained on simulation data to quickly predict complex weld pool outcomes, like geometry and temperature. The study found that GP models provided lower prediction errors than neural networks for all eight outputs, making them highly reliable for real-time use.
Robotic Coupling and Control Loop: A six-axis robotic arm interacts with the simulation by querying the surrogate model at every step. A proportional controller uses this information to adjust the welding power source setting in real-time, simulating how a robot would adapt during a seam weld.
Effective Local-Pressure Parameter: This parameter is used in the thermal model to account for local pressure effects without fully solving complex fluid dynamics. It helps capture thermocapillary transport—how surface tension affects heat flow—which is crucial for accurate prediction of the weld pool shape.

Terminology

Digital Twin — 2× in this paper · explained in 6 episodes

Summary

The gist The UK Space Agency-supported ISPARK project is developing a physics-grounded digital twin framework for structured process de-risking through robotic welding

Framework Architecture

The proposed architecture is organized into four layers with explicit interfaces, allowing the finite-volume model to generate data offline while the surrogate supplies predictions at control-loop rates The layers include:

** Physics model:**

A finite-volume thermal weld-pool model solves enthalpy transport with solid–liquid phase change and is driven by a time-averaged arc heat input, costing of order 102–103 s per operating point Gravity is carried as a design variable for an explicit invariance check and enters the separate hydrostatic-to-capillary pressure ratio, but not the thermal energy equation

Surrogate layer:

Gaussian-process and neural-network emulators are trained on solver sweeps, mapping a five-element process and environment vector to weld-pool geometry, peak temperature and evaporation rate Each GP output has a nominal predictive interval

Robotic coupling:

A six-degree-of-freedom manipulator with inverse kinematics queries the surrogate at each control step, and the controller adjusts the synergic power-source setting

Sensing interface:

The path by which measurements would update model states and parameters is specified, but the loop is not closed against hardware

Reduced Thermal Weld-Pool Model

The model uses an equivalent heat source parameterized by the CMT operating condition and effective local-pressure parameter Key aspects of the formulation include:

Energy Equation:

Specific enthalpy is advected to account explicitly for latent heat across the melting front, and the transported form of the energy equation is then solved Gravity does not enter Eq. (2); its role in the present implementation is limited to the diagnostic in Eq. (8) and to an invariance check in the surrogate design

Convective Transport:

Liquid-metal convection is represented by an enhanced anisotropic thermal conductivity rather than by solving the momentum equations, which approximates the net geometric effect of thermocapillary transport The effective conductivity is defined as keff = ks + (τ κl − ks) fl, where τ κl and τ κ⊥ are directional multipliers

Surface Tension:

The temperature coefficient of surface tension depends on dissolved surfactant content, and its sign can invert according to the Sahoo–Debroy–McNallan formulation The implementation sweeps Cs = 0.002–0.030 wt% rather than fixing a value

Surrogate Fidelity and Speed-up

The surrogate maps a five-variable design—synergic power-source setting, travel speed, gravity level, the effective local-pressure parameter and sulphur concentration—to eight outputs The comparison between Gaussian process (GP) and neural network surrogates shows that the GP gives lower normalised mean absolute error than the neural network for all eight thermal and weld-pool outputs For ten deterministic boundary cases, 92.5 % of 80 solver responses lie within nominal 90 % GP predictive intervals A three-output query takes 249 μs, corresponding to a measured speed-up of 1.6×10 6 relative to the median finite-volume calculation

Model-Based Closed-Loop Study

In a simulated UR5 L-shaped-seam study, a proportional controller adjusts the synergic power-source setting as travel speed decreases near the corner Finite-volume re-evaluation shows that all 12 closed-loop operating points and 9 of 12 open-loop points satisfy an illustrative 1.0–3.0 mm root-side-width band The controller adjusts the synergic power-source setting S in proportion to the error between the target width and the surrogate’s predicted root-side width Wˆ, subject to a rate limit and configured setting bounds

Limitations and Future Work

The study notes several limitations, including that the free surface is flat (rigid-lid), weld reinforcement and root profile are not resolved Droplet detachment, arc-plasma coupling and solidification microstructure are outside scope Future work should assess model predictions through thermal-vacuum and targeted reduced-gravity tests and evaluate the coupled controller through hardware-in-the-loop trials using a robotic arm Hardware testing should prioritise the effects of pressure and thermal boundary conditions Targeted reduced-gravity testing should assess the mechanisms omitted from the model and the relevance of the separate hydrostatic-to-capillary pressure ratio

Conclusions

This paper combines a reduced thermal weld-pool model, data-driven surrogates and a robotic-control simulation for screening candidate in-space arc-welding conditions The thermal outputs are invariant to the sampled gravity coordinate because gravity is absent from the implemented energy equation; no physical gravityinsensitivity claim follows while momentum, free-surface deformation, arc behaviour and droplet transfer remain unresolved Further development requires formal grid convergence, experimental calibration and independent coupon validation Bench trials at 1 g and atmospheric pressure should be conducted to test the pressure-dependent assumptions used by the model Hardware-in-the-loop trials should assess seam tracking, contact-tip-to-work distance, arm and target compliance, command and sensing latency, disturbances, off-nominal events and safe shutdown The paper reports numerical quality control criteria and surrogate fidelity to the finite-volume model The work was supported by the UK Space Agency under National Space Innovation Programme 2 (NSIP 2) through the Intelligent SPace Arc-welding Robotic Kit (ISPARK) project