Efficient learning strategy for predicting glass forming ability in imbalanced datasets of bulk metallic glasses

The prediction of glass forming ability (GFA) and various properties in bulk metallic glasses (BMGs) pose a challenge due to the unique disordered atomic structure in this type of material. Machine learning shows the potential ability to find a way out. However, the training set from the experimental data of BMGs faces the issue of data imbalance, including the distribution of data related to elements, the range of performance data, and the distribution of sparse and dense data area in each specific system. In this work, the origin of the data imbalance and its impact on the GFA prediction ability of machine-learning models are analyzed. We propose the solutions by training the model using the pruned dataset to mitigate the imbalance and by performing an active experimental iterative learning to compensate for the information loss resulting from data reduction. The strategy is proved in Zr-Al-Cu system, and the automated workflow has been established. It effectively avoids the prediction results from trapping into the intensive training-data area or from inducing by the data distribution of similar element systems. This approach will expedite the development of BMGs compositions, especially for unexplored systems.

Figure 1

Distribution analysis of GFA datasets (a) Statistics of element occurrence frequencies in the datasets. (b) Pie chart of label value types. The “Crystal” label represents data that can only be synthesized as crystals, “Ribbon” represents data corresponding to ribbon samples obtained by melt spin quenching, and “BMG” represents data that can be synthesized as BMGs. (c) Statistical chart of $D_{\max }$ of BMGs in the datasets.

Figure 2

Machine learning and experimental iteration workflow proposed in this work. Models are trained and optimized using hyperopt, with model fusion improving prediction performance. The top 5% points exhibiting high GFA are selected based on ranked $D_{\max }$ prediction. Two candidates are experimentally validated in each iteration process and their results are added to the training set for the next iteration.

Figure 3

The impact of varying degrees of data imbalance on the model prediction: (a) diagram depicting the data-type division in a ternary system; (b) diagrams illustrating the data range of the subdatasets; (c)–(g) prediction results of La-Al-C, Co-Ta-B, Ir-Ni-Ta, Mg-Zn-Ca, Zr-Al-Cu, system [$(x-1)\sim (x-5)$ represent the prediction results in the cases of (1) to (5) subdatasets in turn and ($x-6$) represents the true value of the experimental reports, where $x=b,c,d,e$, and $f$]; (h) effect of pruning relevant data on the error of model prediction results (the labeled numbers indicate the data entries reduced from the previous dataset).

Figure 4

Iterative prediction empowered by adding experimental data step by step. (a) reported $D_{\max }$ for Zr-Al-Cu system, where pentagrams correspond to the iteration data points; (b) top 5% iterative prediction results for the Zr-Al-Cu system, showing initial results without iteration (b-1) and after the first (b-2), second (b-3), and third (b-4) iterations.