Session: MF-17-02 Advanced and Additive Manufacturing and Material Technologies (joint with D&A)-2

Paper Number: 153329

153329 - Bayesian Optimisation of a Blown-Powder Additive Laser Process for Pressure Vessel Cladding 

Abstract: 

Cladding for nuclear pressure vessel applications is often necessary to provide a corrosion resistant barrier under pressurised water reactor (PWR) applications. Laying down a stainless steel or nickel-based alloy as the clad on to the pressure vessel substrate (often a low-carbon steel) is performed using an arc-based process such as Tungsten Inert Gas (TIG) or Metal Inert Gas (MIG) process. This is a proven and compliant technique, but requires vessel pre-heat, interstage non-destructive examination (NDE) and post-clad machining, all of which make the cladding process expensive and lengthy.

Trials have demonstrated that there is potential to significantly reduce the cladding (bulk cladding) process time, cost and carbon emissions by adopting a blown-powder laser-based approach. In this additive process, a laser is used to fuse a metal powder which is fed into[HD1] [GN2]  a laser beam whose focal point is near the surface of the substrate. The result is a single-pass, thin-layer clad which may not require post-clad machining or inter-stage NDE.

The process is governed by a number of key parameters, including the laser power (typically between 3-15 kW[HD3] ), transition speed (rotation speed of bore with respect to laser in 2G position), powder feed rate and clad offset. During a process development programme, these key parameters are tuned in order to meet certain acceptance criteria, including geometrical, roughness and chemical composition criteria. Finding the right set of parameters is extremely difficult because the total number of possible process parameter combinations is very high, resulting in lengthy technique maturity programmes[HD4] [GN5] .

Bayesian Optimisation is well-suited to developing and optimising advanced manufacturing processes, as it lends itself well both to cases where the objective function is expensive to evaluate (it doesn’t require large numbers of process runs) and effectively handles multiple output constraints.

This paper reports on the use of a Bayesian Optimisation technique to tune the process parameters in a series of blown-powder additive manufacturing trials.[HD6]  This follows a desktop study presented previously at PVP (PVP2023-106320) which aimed to show how Bayesian Optimisation could[HD7] [GN8] [HD9] [HD10]  be deployed to a complex advanced manufacturing process such as laser cladding, reducing the guesswork and trial-and-error approach previously adopted, and in doing so, drastically reduce number of manufacturing trials required, and therefore the time and cost of a technique maturity programme.

As part of the Bayesian Optimisation process presented in this paper, an acquisition function is defined to simultaneously account for the multiple acceptance criteria, while also optimising the process speed. The method has also been developed to account for robustness criteria in a natural probabilistic manner, so as to select process parameters that reduce the risk of non-conformity. We show how this technique is used to more efficiently tune the process parameters to achieve an acceptable result, saving time and cost in the technique development programme.

Presenting Author: Greg Nelson Frazer-Nash Consultancy Ltd.

Presenting Author Biography: Greg is a chartered Mechanical Engineer with a strong background in structural analysis and design. He has been working at incorporating probabilistic and machine learning methods with physics-based techniques across a range of sectors. He has been developing and applying Bayesian Optimisation methods in the nuclear (fission and fusion) design and manufacturing sectors, and his wider structural analysis experience includes assessment of fault scenarios, fatigue and fracture mechanics and static and dynamic simulation.

Authors:

Greg Nelson Frazer-Nash Consultancy Ltd.
Steve Lawler Frazer-Nash Consultancy Ltd.