1 7 A D V A N C E D M A T E R I A L S & P R O C E S S E S | A P R I L 2 0 2 3 dimensional problems (e.g., 4 dimensions and above), but for illustrative purposes only three- dimensional space is easiest to imagine. In the conventional approach, many samples are collected at only one parameter point (illustrated by the red dot in Fig. 2a). In the ML approach (illustrated by the graph in Fig. 2b), the ML software was used to design the design of experiments (DOE) that consist of each of these blue dots. Each of these blue dots was built using a different parameter set. Each of the blue dots was used to develop a surrogate model (sometimes called a response surface) of the AM process. The ML model can point to parameter A and parameter B even though no empirical data has been collected at these dots. Each of these blue dots is not just a point but is a processing window. The ML approach enables the user to conduct parameter development and allowables development simultaneously. In other words, the blue dots inform the parameter optimization and allow the AM user to select the optimal parameters to achieve the goals, but the blue dots also allow the AM user to make statistically substantiated predictions about the performance and the coefficient of variation of those predictions at any given point or process window. Hence, ML allowables could be calculated at the red dot for Goal A and done a second time at the pink dot for Goal B. In comparing ML allowables against conventionally developed MMPDS or CMH-17 ones, the ML allowables predicted materials behavior consistently with the conventionally developed ones. In other words, ML allowables are just as accurate as those conventionally developed. Table 3 presents an example comparing ML allowables against conventional MMPDS S-basis allowables for parameter A and B from the AMMP project on stainless steel 17-4 PH. It is critical to recognize that the ML approach enables a user to do parameter development and materials allowables development in parallel using the exact same empirical dataset. Simultaneously, the single set of training data that is gathered in the ML approach can be used to generate an infinite number of allowables, thus the cost-benefit becomes even more favorable toward the ML approach if more than one allowable is generated. Figure 3 provides a chart showing the total cost of allowable development (y-axis) for various quantities of allowables. For the purposes of this chart, the conventional MMPDS S-basis allowables were assumed to be developed using the minimum quantity of coupons required as per MMPDS (i.e., 30 coupons). MATURATION OF ML FOR AM ALLOWABLES DEVELOPMENT The future state of AM of mechanical property allowables development and AM process qualification is likely to involve the integrated use of the following technologies. • Integrated computational materials engineering (ICME) • Machine learning (ML) and artificial intelligence (AI) Fig. 2 — (a) Conventional allowable development; X (e.g., 30) samples at one parameter set (e.g., A1, B1, C1), so all 30 samples are repeats of the same parameter set; (b) ML approach; Y samples all over and evenly distributed over the parameter space; S-basis “ML-allowables” can be calculated at any parameter set even where there’s no empirical data (e.g., red, pink dots). TABLE 3 — CALCULATED MACHINE LEARNING ALLOWABLES RESULTS COMPARED TO CONVENTIONAL APPROACH PARAMETER A — OPTIMIZED TO PRIORITIZE HIGH TENSILE STRENGTH OVER PRINT SPEED ML allowable (MPa) MMPDS allowables (MPa) Ultimate tensile strength 154.9 159.4 Yield strength 153.0 153.2 PARAMETER B — OPTIMIZED TO BALANCE BETWEEN GOOD TENSILE STRENGTH AND FASTER PRINT SPEED Ultimate tensile strength 160.6 164.4 Yield strength 156.5 157.1 (a) (b)
RkJQdWJsaXNoZXIy MTYyMzk3NQ==