Objective: Based on the author´s four-decades long global tribological experience, approaches to friction material compounding (or formulating) may be broadly categorized as: (a) informed one-ingredient or few-at-a-time change approach, (b) statistical- or DoE (Design of Experiments)-based approach, and (c) AI-/ML-based approach combining approach (b) with an ML (Machine-Learning) method. In the author´s experience, approaches (a) and (b) have been demonstrated to be both time-consuming and inefficient for accuracy when compared with approach (c). But why does this matter? 1) Ever shorter vehicle development cycles – including the advent of EVs – with ever more stringent brake pad/lining technical requirements, which have to be delivered with ... 2) A relatively new Copper-free friction material development base, while keeping in mind that ... 3) The human mind is fundamentally unable to analyze very large-scale systematic data and find trends to meaningfully extract forward- or backward-propagating predictive insights. The above three together establish AI-/ML-based Compounding as a key need for Brake Industry 4.0 to deliver fast innovation and design-for-quality while enabling a lower-cost efficient “idea-to-market” process. Methodology: While the actual ML-based methodology used will not be the subject of this paper (given that each business entity may pursue its own proprietary ML method), the emphasis herein is on the overall I/O (input-output) definition with forward- / backward-propagation predictive modes, the creation of a “clean” platform database, and the mandatory deployment of friction material-specific rules (with relevant underpinnings in Composite Materials theory). As a case in point, a rudimentarily-trained forward-propagating ML-based tool, absent any such rules, may logically predict a “low-fade” formulation as one almost devoid of organics including resin; this may, however, result in a structural catastrophe. Permitting the introduction of prior “expert human learnings” in the form of a few rules may be an elegant way of avoiding an utterly meaningless output from an ML-based compounding tool. Similar sensible rules for other functional ingredients and production process parameters, plus drivers for optimum ratios between the number of “clean” input data and output variables, all aimed at delivering a first-cut reasonable prediction accuracy, will also be presented. Results: AI-/ML-based formulations based on forward- and backward-propagating predictions deploying these rules yielded dynamometer tested accuracies no worse than standard test-to-test variability shown by multiple pads from one lab-made batch. Formulations so developed were launched in on-road vehicle applications. The well-trained ML-based tool also yielded critical raw material and process parameter Sensitivity Graphs which enabled a manufacturing robustness geared towards reduced piece-to-piece variability. Rules and some examples from this array of results may be presented in a generic manner. Limitations: The encountered limitations are indeed quite simply opportunities for future and, in some cases, possibly collaborative work. This may include but are not limited to: (a) meaningful combinations of platforms (front-/rear-axle) that can benefit from one ML database, (b) ML modules for wear, NVH, static mu, corrosion, coated discs, relaxation, particle emissions, etc. What is new: As AI-/ML-based compounding efforts are relatively new, the purpose of this paper is to leverage the author´s prior experience in AI-/ML-based compounding from about three decades ago to help shorten the learning-curve for the enthusiastic newcomers of today and ensure that they create as few pitfalls for themselves along the product development and launch processes. The rapidly burgeoning EV introduction across wide swathes of vehicle platforms requires that the friction industry be in lock-step with not only efficient and cost-effective product development tools but also with potentially the same tools that can deliver an ability to data-mine useful information from past know-how / development efforts and build in quality aspects to deliver manufactured piece-to-piece consistency. Conclusions: The necessity for strict and sensible accompanying rules for AI-/ML-based Friction Materials Compounding is demonstrated. The general nature of these rules is expounded upon. The author studiously opines that this early-generation AI-/ML-based approach is meant to be a useful tool in the hands of a disciplined friction material compounder, and not a means to replace the compounder. This enhanced human expert-machine partnership – working in unison as an effective DeskTopDynoTM – will have the ability to deliver better predictive capabilities than either one while simultaneously delivering a studied basis for designed-in quality. The ML-based tool also lends itself as an effective training means.

Mr. Parimal Mody, Automotive Brake and Friction Materials Expert, PBM Consulting