PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Zhou Daniel Hao, Umer Masood Chaudry, Robert Shaw, Andrew Cheney, Nick Ludford, Mark Sims, Hongbiao Dong

arXiv:2610.21939v1cs.ROphysics.app-ph

TL;DR

In-space robotic welding remains poorly characterised under reduced gravity and low pressure, while flight-like tests are limited. This paper combines a finite-volume weld-pool model, surrogates, and simulated robotic control; the GP outperforms the neural network across outputs, and closed-loop stations meet the illustrative width band more often than open-loop stations.

  • Problem

    Metal joining remains largely unqualified for flight, and existing studies do not establish a welding procedure or operating envelope across mission-relevant gravity and pressure conditions.

    Source: §1 · p. 2

  • Method

    The paper combines a reduced finite-volume thermal weld-pool model, GP and neural-network surrogates, and a simulated six-axis manipulator for model-based control.

    Source: §1 · p. 2, §11 · p. 13

  • Results

    All 12 closed-loop operating points met the illustrative root-side-width band, versus 9 of 12 open-loop points; the GP was more accurate than the neural network for all eight outputs.

    Source: §11 · p. 13

  • Takeaways & Limitations

    The framework supports screening candidate in-space arc-welding conditions and control strategies before experimental tests.

    Source: §11 · p. 13

    The model uses a flat rigid-lid free surface and does not resolve weld reinforcement, root profile, droplet detachment, or arc-plasma coupling.

    Source: §3.6 · p. 6

Abstract

from arXiv · show

Growing interest in in-space servicing, manufacturing, and assembly has renewed attention on reliable metal joining for modular construction and on-orbit repair. Robotic welding is promising, but reduced-gravity and low-pressure operation remains poorly characterised and flight-like tests are limited. The UK Space Agency-supported ISPARK project is developing a physics-grounded digital twin for structured process de-risking. This paper evaluates a model-based prototype combining a finite-volume thermal weld-pool model, Gaussian-process (GP) and neural-network surrogates, and a simulated six-axis manipulator. The reduced model includes solid-liquid phase change, effective thermocapillary transport, and a local-pressure parameter. Gravity is assessed separately through a hydrostatic-to-capillary pressure ratio and does not enter the thermal energy equation. Across repeated five-fold cross-validation, the GP has lower normalised mean absolute error than the neural network for all eight outputs. In 10 held-out boundary cases, 74 of 80 solver responses fall within nominal 90% GP predictive intervals. A three-output GP query takes 249 microseconds, including scaling and predictive variance. In a simulated UR5 L-shaped-seam study, a proportional controller adjusts the synergic power-source setting as travel speed falls near the corner. Finite-volume re-evaluation finds all 12 closed-loop operating points and 9 of 12 open-loop points within an illustrative 1.0-3.0 mm root-side-width band. These simulation results support using the framework to screen welding conditions and control strategies before experimental tests. Thermal-vacuum, targeted reduced-gravity, and robotic hardware-in-the-loop studies are needed to assess the model and controller.

1. Introduction

In-space metal joining remains largely unqualified for flight, and existing studies do not establish a procedure or operating envelope across mission-relevant gravity and pressure conditions. This paper proposes a physics-grounded model and surrogate framework for screening conditions and control strategies before hardware tests.

1. Introduction

  • Existing space-welding studies do not report a qualified procedure or operating envelope spanning mission-relevant gravity and pressure conditions.Experiments are resource-intensive, and each covers only a small part of the parameter space.Source: §1 · p. 2
  • The study applies a finite-volume thermal model to 1 mm 316L sheet, with explicit numerical convergence and field-quality criteria; the 2 mm coupon configuration is not evaluated.A formal grid-convergence study and experimental validation against ISPARK coupons remain unavailable.Source: §1 · p. 2
  • GP and neural-network surrogates are evaluated by repeated cross-validation and deterministic boundary cases, including predictive-interval coverage and inference latency.The finite-volume model is too costly for robotic control rates or dense exploration of a five-dimensional design space.Source: §1 · p. 2
  • A model-based closed-loop study tests a controller that adjusts the synergic power-source setting in response to trajectory-imposed changes in travel speed.The study couples the GP surrogate to a six-degree-of-freedom manipulator simulation.Source: §1 · p. 2
  • The sensitivity analysis separates conclusions supported within the reduced model from gravity-dependent free-surface and metal-transfer mechanisms that still require experiments and higher-fidelity models.Gravity is sampled for numerical invariance and enters a separate hydrostatic-to-capillary pressure ratio, but not the implemented thermal energy equation.Source: §1 · p. 2

2. The ISPARK digital-twin architecture

The proposed architecture separates offline finite-volume training-data generation from online surrogate predictions for a simulated robotic control loop. Its sensing interface is specified but not tested with hardware, so the implementation is a model-based prototype rather than a hardware-connected digital twin.

2. The ISPARK digital-twin architecture

  • The finite-volume weld-pool model generates training data offline, while Gaussian-process and neural-network surrogates serve predictions to the robotic control loop online.The architecture separates the physics model, surrogate, robotic coupling, and sensing interface.Source: §2 · p. 3, Figure 2 · p. 3
  • The thermal model includes solid–liquid phase change and effective local-pressure effects, while gravity is assessed separately and does not enter its energy equation.Gravity enters a separate hydrostatic-to-capillary pressure ratio; the model costs about 10^2–10^3 seconds per operating point.Source: §2 · p. 3
  • A simulated six-degree-of-freedom manipulator follows the seam and queries the surrogate at each control step, with the controller adjusting the synergic power-source setting.Travel speed is prescribed by the motion plan.Source: §2 · p. 3
  • The sensing interface and assimilation structure are specified but not implemented, making the current system a model-based control prototype rather than a hardware-connected digital twin.Transfer to laser- or electron-beam welding is not automatic and would require process-specific modeling, inputs, training data, and validation.Source: §2 · p. 3

3. Reduced thermal weld-pool model

The reduced weld-pool model combines enthalpy-based phase change, effective thermocapillary transport, and pressure-sensitive surface losses to screen welding conditions. Its simplifications constrain what it can predict, particularly fluid motion, free-surface shape, and low-pressure effects.

3.1 Formulation

  • The model transports enthalpy with latent heat across the solid, mushy, and liquid states, using a parameterised heat source rather than resolving the arc, electrode, or metal transfer.The reported model targets full-penetration welding of unsupported 1 mm sheet; gravity is assessed separately and does not enter the energy equation.Source: §3.1 · pp. 3–4, §3.5 · p. 5

3.2 Convective transport as an anisotropic effective conductivity

  • Liquid-metal convection is approximated by anisotropic conductivity: its direction follows the surface-tension temperature slope, but it does not predict velocity, recirculation, or mixing time.The fixed enhancement factors are not an independently calibrated decomposition, and sign changes within a pool can produce unresolved counter-rotating regions.Source: §3.2 · p. 4

3.3 Sulphur-dependent surface tension

  • Sulphur concentration is swept because it can change the sign of the surface-tension temperature coefficient, altering predicted flow direction and weld-pool geometry.The study uses 0.002–0.030 wt% sulphur; the transition may occur within the weld-pool temperature range, where the effective-conductivity model cannot resolve counter-rotating flow.Source: §3.3 · pp. 4–5

3.4 Reduced-pressure evaporation

  • The pressure-sensitive surface-loss model uses an imposed local far-field pressure, but its evaporation coefficients are uncalibrated and that parameter is not automatically chamber pressure.The gas jet, arc plasma, vapour plume, and pressure-dependent arc constriction are unresolved, so mapping chamber pressure and gas flow to the parameter requires separate modelling or measurement.Source: §3.4 · p. 5

3.5 Thin-sheet boundary conditions

  • For unsupported 1 mm sheet, both faces are treated as free surfaces and a finite-fin boundary represents heat loss through the unmeshed strip toward the fixture.The hydrostatic-to-capillary ratio excludes arc pressure, free-surface deformation, and dynamic metal transfer, so it is not a standalone burn-through predictor.Source: §3.5 · pp. 5–6

3.6 Limitations

  • The rigid-lid, time-averaged model omits free-surface shape, controlled dip-transfer or pulsed-current behaviour, droplet detachment, arc-plasma coupling, and solidification microstructure.Higher-fidelity work would require linked thermal, fluid-flow, interfacial, and solidification models supported by experimental inputs and validation evidence.Source: §3.6 · p. 6

4. Numerical convergence and quality control

The solver uses simultaneous drift, storage, energy-balance, and local-field checks to judge convergence. The retained results meet those criteria, but formal verification remains incomplete and experimental validation is absent.

4. Numerical convergence and quality control

  • The work reports numerical quality-control checks, not experimental validation, and a formal grid-convergence study remains outstanding.Verification concerns numerical solution of the governing equations; validation concerns agreement with physical observations.Source: §4 · p. 6

4.1 Convergence criterion

  • Convergence requires four simultaneous checks: drift, stored-enthalpy rate, energy-balance closure, and local temperature-field quality.The criteria include less than 2×10^-3 change across six checks, storage below 4% of net input power, energy closure within 8%, and local field-shape constraints.Source: §4.1 · p. 6

4.2 Outstanding verification studies

  • 68 retained results meet the stated convergence and field-quality criteria, but formal numerical verification remains incomplete.Outstanding studies include grid refinement, domain and initial-condition independence, and a tabulated steady-state energy balance; predictive intervals therefore quantify surrogate uncertainty only.Source: §4.2 · p. 7

5. Literature-informed parameter priors

The study uses literature-informed priors because it found no directly matching experimental dataset for the target material, thickness, and operating range. The parameters are not calibrated to ISPARK measurements, and several evidence gaps limit interpretation.

5. Literature-informed parameter priors

  • No direct validation dataset was identified for 1–2 mm 316L over the present travel-speed range, so uncertain parameters use literature-informed priors rather than ISPARK fits.The study specifies priors only; it reports no posterior calibration or model-discrepancy estimate.Source: §5 · p. 7
  • The priors cover arc efficiency, sulphur-dependent surface tension, heat-source distribution and droplet power, and effective-conductivity factors for unresolved convection.Arc efficiency spans 0.65–0.87; heat-flux and current-density prior scales are 1–3 mm.Source: §5 · p. 7
  • No applicable burn-through threshold was identified for unsupported thin stainless-steel sheet, and low-pressure evaporation coefficients for 316L were unavailable.Evaporation is therefore treated as a sensitivity rather than a validated prediction.Source: §5 · p. 7

6. Surrogate modelling and benchmark

The paper builds GP and neural-network surrogates over a five-variable design and evaluates their fidelity to a finite-volume model. It also tests whether scalar surrogate outputs can reconstruct weld-pool shape.

6.1 Design and construction

  • The five-variable design maps power-source setting, travel speed, gravity, local pressure, and sulphur concentration to eight weld-pool outputs.Cases outside the nominal net-heat-input range were filtered; gravity is included for a numerical invariance check but is inactive in the thermal equations.Source: §6.1 · p. 7
  • A Matérn GP and a two-hidden-layer neural network are compared using repeated cross-validation and held-out deterministic boundary cases.The GP is fitted independently by output and supplies output-specific predictive intervals; cases with numerical failures or active safeguards are excluded from training.Source: §6.1 · pp. 7–8

6.2 Prediction accuracy and computational cost

  • Across 80 boundary-case responses, 74 fall within nominal 90% GP intervals, but coverage varies from 60% for root-side width to 100% for six outputs.The reported coverage does not establish calibrated uncertainty for every output.Source: §6.2 · p. 8

6.3 Weld-pool shape reconstruction from scalar outputs

  • At the reference condition, a transverse trapezoid formed from top width, root-side width, and depth achieves IoU = 0.93 against the finite-volume liquidus boundary.This result is specific to the reported 1 mm case, whose pool walls are nearly vertical.Source: §6.3 · p. 8
  • In plan view, axial placement raises the length–width ellipse’s IoU from 0.68 at the torch to 0.95, because position dominates the error.The contour extends about 0.7 mm ahead of the torch; a leading-edge offset would require regenerating the training sweep.Source: §6.3 · p. 8

7. Closed-loop robotic coupling

A simulated UR5 seam study uses a proportional controller to adjust the power-source setting as travel speed changes near a corner. Finite-volume re-evaluation places all closed-loop stations, but not all open-loop stations, within the illustrative width band.

7.1 Computational requirements for control

  • The offline finite-volume model requires about 10^2–10^3 seconds per operating point, making direct control impractical; evaluating 217 seam steps would take approximately 25–28 hours.The model is retained as an offline reference rather than used directly in the control loop.Source: §7.1 · p. 8

7.2 Implementation

  • The simulated UR5 follows a 50 mm L-seam, where a corner-speed reduction raises heat input per unit length by one third at fixed machine settings.The controller changes the synergic power-source setting while travel speed remains prescribed by the motion plan.Source: §7.2 · pp. 8–9
  • The proportional controller targets 2.0 mm root-side width with gain 0.28, rate-limiting each setting change; its local algebraic error multipliers are 0.17 and 0.47.These values satisfy the stated stability condition for the unsaturated algebraic model, not for a future transient physical system.Source: §7.2 · p. 9

7.3 Scope

  • The surrogate evaluation and finite-volume cross-check remain model-to-model evidence; the simulation excludes robot–structure dynamics and does not establish performance against an independent physical system.The cross-check reevaluates commanded operating points at steady state and does not resolve thermal lag.Source: §7.3 · p. 10, §7.4 · p. 10

7.4 Finite-volume-model assessment of commanded operating points

  • All 12 closed-loop stations fall within the illustrative root-side-width band, versus 9 of 12 open-loop stations, in finite-volume re-evaluation.The maximum open-loop excursion is 0.678 mm; stations are clustered near the corner, so counts are not seam-length fractions.Source: §7.4 · p. 10, Table 1 · p. 10

8. Model-based sensitivity and transfer limits

The model-based sensitivity diagnostics identify power-source setting and travel speed as the strongest influences on predicted weld-pool geometry, while gravity is inactive in the thermal equations. The simulated corner controller limits the modeled heat-input rise, but the results do not establish physical gravity insensitivity or transfer across welding processes.

Model-based sensitivity and transfer limits

  • Power-source setting and travel speed have the shortest GP length scales and largest predicted influence on weld-pool dimensions; gravity’s predicted penetration-depth change is less than 10−4% of its sampled output range.Gravity is inactive in the thermal equations, so this is a numerical invariance check rather than a physical sensitivity result.Source: §8 · p. 10
  • The imposed local-pressure parameter acts most strongly on modeled evaporation through assumed surface-loss relations, while sulphur concentration influences surface-tension-dependent geometric responses.The diagnostic describes the reduced model, not experimental process sensitivity.Source: §8 · pp. 10–11
  • A speed drop from 12.0 to 8.94 mm s−1 raises heat input per unit length by 34% at fixed setting, while control limits the rise to 5.5% and keeps the surrogate prediction inside the band.The L-shaped-seam study assumes a microgravity/vacuum reference condition for 1 mm 316L and uses a UR5 cell.Source: §8 · p. 11, Figure 5 · p. 11
  • The model excludes momentum, free-surface deformation, gravity-dependent droplet detachment, and arc behaviour, so its thermal-output invariance does not establish physical gravity insensitivity.Transfer to GTAW, other GMAW transfer modes, laser welding, or electron-beam welding requires process-specific heat-source models, data, and validation.Source: §8 · p. 11

9. Sensing interfaces

The framework is designed to accept measurement updates, but this study does not test the sensing interface with hardware. Proposed channels support between-weld parameter updates; within-weld estimation requires additional models and an estimator.

Sensing interfaces

  • Candidate sensing channels include power-source current and voltage, calibrated thermal imaging, and weld-pool and bead vision; a root-side view could measure thin-sheet root width and burn-through onset.Current and voltage can be acquired at kilohertz rates; thermal imaging requires radiometric calibration, an emissivity model, and treatment of arc radiation.Source: §9 · p. 11
  • In the Kennedy–O’Hagan structure, each channel would enter as an observation with its own error model to update parameter posteriors between welds; within-weld estimation additionally needs an observation model, estimator, and transient process model.The measurement interface was not tested with hardware, and forward query latency alone does not establish within-weld capability.Source: §9 · p. 11
  • Synchronised sensor records, model versions, and parameter estimates can provide traceability, but inspection and acceptance remain outside the model-based study.Source: §9 · p. 11

10. Future testing and model development

The paper calls for experimental assessment of predictive accuracy and the coupled robotic system, while distinguishing validation from weld-procedure qualification. Proposed work prioritizes thermal-vacuum, targeted reduced-gravity, and hardware-in-the-loop testing alongside model extensions.

Future testing and model development

  • The present study is limited to modelling and simulation; proposed experiments would assess predictive accuracy and robotic-system performance, while procedure qualification requires separate structural acceptance criteria and applicable welding standards.The illustrative root-side-width band is not a qualification criterion.Source: §10 · p. 11
  • Figure 6 gives reduced-model GP sensitivity diagnostics, while Figure 7 shows modeled liquidus-contour dimensions of 5.2 mm pool length, 2.50 mm top width, and 2.00 mm root-side width.The Figure 7 dimensions are modelled liquidus-contour dimensions, not measurements of a solidified weld bead; Figure 6 is not experimental process sensitivity.Source: Figure 7 · p. 12, Figure 6 · p. 12
  • Thermal-vacuum tests should vary chamber pressure, radiative-sink temperature, and initial coupon and fixture temperatures to test pressure- and temperature-dependent assumptions.Proposed measurements include heat rejection, local gas pressure and composition, evaporation-related mass loss, arc behaviour, penetration, and bead geometry.Source: §10 · pp. 12–13
  • Targeted reduced-gravity tests should assess free-surface deformation, metal transfer, and weld-pool and final-bead geometry, which the thermal model does not resolve.Thermal-vacuum testing is identified as the immediate priority for assessing the pressure-dependent assumptions.Source: §10 · p. 13
  • Hardware-in-the-loop trials should connect the deployed surrogate and controller to a manipulator, welding end-effector, sensing chain, and representative fixture, assessing tracking, latency, disturbances, off-nominal events, and safe shutdown.The proposed progression runs from non-welding motion and sensor emulation to integrated welding in a thermal-vacuum environment.Source: §10 · p. 13
  • Model extensions should prioritize free-surface deformation and address droplet detachment, the CMT current and wire-motion cycle, robot–structure dynamics, and solidification microstructure; each requires numerical verification and experimental validation.Test records should link model versions, parameter estimates, sensor measurements, and weld observations to trace changes in predictive performance.Source: §10 · p. 13

11. Conclusions

The paper combines a reduced thermal weld-pool model, surrogates, and robotic-control simulation to screen candidate in-space arc-welding conditions. The GP outperforms the neural network across eight outputs, and all 12 simulated closed-loop points meet the illustrative width band; physical weld predictions remain unvalidated.

Conclusions

  • The GP is more accurate than the neural network for all eight outputs, and 92.5% of finite-volume responses in ten deterministic boundary cases fall within nominal 90% GP intervals.A three-output query takes 249 μs, a measured 1.6×10^6 speed-up relative to the median finite-volume calculation.Source: §11 · p. 13
  • All 12 closed-loop L-seam operating points meet the illustrative root-side-width band, compared with 9 of 12 open-loop points.This supports feasibility of the simulated control strategy but does not validate predictions for physical welds.Source: §11 · p. 13
  • Power-source setting and travel speed dominate modeled geometric responses, while local pressure affects predicted evaporation; gravity invariance does not establish physical gravity insensitivity.Further work includes grid convergence, experimental calibration and independent coupon validation, thermal-vacuum and reduced-gravity tests, HIL evaluation, and transient free-surface and metal-transfer extensions.Source: §11 · p. 13
Loading 2610.21939v1…