Challenge 2: James Walker

James Walker: Machine Learning Approaches for Predicting Optimum Performance of Advanced Elastomers

James Walker is an engineering and technology company. Through their Materials Technology Centre in the UK, James Walker develops and compounds its own range of advanced elastomer materials, holding over 300 unique formulations including many bespoke developments for specific customer applications across the biopharmaceutical, nuclear, defence, and oil & gas industries.

Elastomeric materials can have surprisingly complex material formulations - because of this, development of new compounds is often a lengthy trial-and-error process (sometimes described as a ‘dark art’). Most manufacturers do tend to rely on families of elastomeric materials where their performance is controlled by the underlying reinforcement i.e. (nano)particle (e.g. carbon black) with different level of loading, particle sizes or changes to the elastomer chemistry (e.g. cross-linking density). This leads to a large number of design variables that requires extensive testing to meet a given standard (either customer- or international regulation body-specific) which frequently cannot be realised due to time and economic restrictions. Ultimately, the elastomer/seal material industry would like to have predictive models that reliably model seal life expectancy which is the holy grail of that industry.

Therefore, the ultimate challenge here is to propose a range of predictive approaches for improving the quantification of failure risk depending on the material composition (e.g. different filler contents and/or chemistries) without excessive and costly testing.

Specifically, the challenge is to explore if available machine learning approaches (e.g. Bayesian framework) can:

• Capture non-linear large deformation response of advanced elastomers using experimental data and their probabilistic description;
• account for uncertainties arising from incomplete experimental data (in some cases) and from variations in responses/behaviour for different sample batches;
• quantify the risk of failure according to safety criteria used in the industry, and use probabilistic approach to identify optimum elastomer compositions that would ensure safe operation of a material/component.

Approaches

The participants will be encouraged to consider machine learning that can capture uncertainties (e.g. Bayesian framework) to develop surrogate descriptions of material behaviour observed experimentally.

James Walker Ltd will provide research context into the topic and guidance on the specific material composition of interest, along with experimental data for selected

industrially-relevant systems, including two distinct families of elastomeric materials, fluoroelastomer (FKM) and acrylonitrile-butadiene (NBR), both reinforced with carbon black and utilising different cross-linking systems).

The initial data sets will be based on unaged materials. In the longer term Jams Walker would like to be able to use the developed models/approaches with aged data sets allowing us to reliably model seal life expectancy, and adjust to the requirements of different industries that tend to use different criteria to determine end of life - so any model/approach should be flexible enough of taking this into account.

Extra information

Traditionally elastomer behaviour (including end of life) and their material selection by either time-consuming trial-and-error experimentation or estimated through deterministic phenomenological constitutive modelling and deterministic safety factors for material behaviour and structures. However, for in-service behaviour predictions the latter also become computationally, especially in the multiscale modelling context, when finite element solvers need to be concurrently linked across the scales (e.g. material microstructure scale with the structural response). Machine learning approaches, including Bayesian, hold the promise of making those calculations more efficient, and enable accounting for uncertainty (Bayesian). While a complete risk assessment would require 3D data (and tensorial description), this challenge will provide a proof-of-concept that machine learning can help in risk assessment and optimisation of elastomer behaviour within a lower dimensional setting, and probabilistic end of life risk prediction.