Page 141 - Read Online

Chen et al. Energy Mater. 2025, 5, 500120 https://dx.doi.org/10.20517/energymater.2024.311 Page 17 of 21

Figure 9. Showing the iterative augmented strategy to understand the melt pool geometrical (width and depth) using LPBF processing
of n-type Bi Te Se thermoelectric. A total of six iterations were performed with feedback [63] . (License number 6015220339088).
2
0.3
2.7
[65]
is by Na et al. . This study converts chemical reaction formulas into a machine-readable format using a
"synthesis graph". The synthesis graph describes the chemical formula in terms of the elements present in
the starting and final materials, as shown in Figure 10. To predict the synthesis recipe of TE materials, a
DNN-based architecture called the synthesis process encoder-decoder (SPENDE) has been developed
[Figure 11]. Herein, based on the benchmark of synthesis dataset of 771 unique TE materials; first, a
synthesis graph has been generated followed by calculation of graph embedding vector using a graph NN
graph-based reaction encoder (GRE). Then, the NN operation sequence decoder (OSD) has achieved the
prediction of each step's operation level. Finally, the preparation conditions are predicted by engineering
conditions networks (ECNs) predicts. Thus, this architecture has successfully predicted the synthesis
parameters involving grinding, heating, cooling, and sintering for synthesizing TE materials.

CHALLENGES AND PERSPECTIVES
The rapid development of statistical methods has significantly influenced the discovery and design of TE
material. This critically means that exciting new materials are being discovered; however, these materials are
yet to be produced. ML and AI have the potential to learn patterns of synthesis design from a given data set
of experimental synthesis procedures and then predict the outcome. Nevertheless, the design of synthesis
parameters for TE materials is challenging because the sequence of reactions during the synthesis depends
on many factors including choice of synthesis procedure and precursors. Therefore, the unavailability of an
organized and comprehensive database of synthesis procedures of TE materials is a big challenge that needs
to be overcome. For training the algorithm, converting available synthesis condition data and processing it
into a suitable format for the algorithm is another obstacle. Finally, the synthesis parameters predicted by
statistical optimization can be confirmed by experiments and further improved and developed.

Atomic properties are building blocks to construct the crystal system. Also, these parameters are related to
the TE performance. Experimentally, finding this relation consumes more time and material cost. However,
ML easily and quickly interprets this relation through the available data from materials projects or research
articles. Interpretation of atomic properties with TE parameters tends to material classification which could
help to identify new materials or compositions without any cost. Besides, the ML approach quickly