Page 9 - Read Online

Li et al. J Mater Inf 2024;4:27 https://dx.doi.org/10.20517/jmi.2024.44 Page 3 of 20

optimizing the compositions and heat treatment conditions [21,27] . After testing, it was found that the tensile
properties and creep resistance of CNAs were greatly improved compared with EUROFER 97. Although
CALPHAD can predict microstructures from compositions and heat treatment parameters, it struggles to
establish quantitative relationships with properties, hindering efficient material design.

In recent years, machine learning (ML) method has been applied to discover advanced materials. The core
of this method lies in developing prediction models describing the relationships between compositions,
processing parameters, microstructures, and properties . Once accurate prediction models are established,
[28]
the properties of thousands of candidate materials can be efficiently calculated, guiding subsequent
experimental testing. For example, Wen et al. developed a ML-based hardness prediction model to
[29]
accelerate the discovery of Al-Co-Cr-Cu-Fe-Ni alloys with enhanced hardness . Similarly, Yu et al.
proposed a new design strategy to screen Co-base superalloys with excellent properties from 363,000
candidates using four ML-based prediction models . However, most ML studies mainly focus on
[30]
performance optimization, as limited experimental data poses challenges to developing accurate
microstructural models. Moreover, there are few reports on ML-based design of RAFM steels, aside from
the study by Wang et al. and our recent work [31,32] . A ML model was developed by Wang et al. to predict the
tensile properties of RAFM steels based on their compositions and heat treatments . In our recent work ,
[32]
[31]
we introduced a ML-based intelligent design model to guide the compositional and processing design of
RAFM steels. This model facilitated the development of a new RAFM steel achieving a UTS ~100-400 MPa
higher than that of conventional RAFM steels, with comparable total elongation (TE). Unfortunately, its
[21]
calculated V of ~0.03% is significantly lower than that of FM steels (such as ~0.35% in Grade 91 ). If the
MX
prediction problem of microstructures in this intelligent design model can be solved, it will help to quickly
discover RAFM steels with high V and excellent tensile properties.
MX
In this work, a combination of CALPHAD and ML has been naturally proposed to achieve the integrated
design of structure and performance. Firstly, accurate microstructural dataset was provided by high-
throughput CALPHAD to support ML modeling. Based on this dataset, a microstructural model is
constructed, taking compositions and heat treatments as inputs and microstructural attributes as outputs.
Secondly, an integrated design model is built by combining the microstructural model with the forward and
[32]
reverse models from our former work , executing the structure-property-oriented compositional and
processing design. Finally, this model is employed to develop new RAFM steels with desired
microstructures and tensile properties, followed by necessary experimental validations.

MATERIALS AND METHODS
Design strategy
An accelerated design strategy, incorporating ML and CALPHAD, was applied to develop new RAFM steels
with targeted microstructures and tensile properties, as shown in Figure 1. The process involves the
following steps:

First, three data-driven prediction models were established: a forward model, a reverse model, and a
microstructural model. The forward model predicts the tensile properties of RAFM steels based on their
compositions and processing parameters, while the reverse model suggests candidate combinations of
compositions and processing parameters to achieve targeted tensile properties. The forward and reverse
models with high accuracy were developed and validated in our previous work . The microstructural
[32]
model includes four sub-models to predict: (1) the presence of δ-ferrite at normalizing temperature (NT);
(2) the presence of coarsening phases (i.e., Laves and Z-phase) at tempering temperature (TT); (3) the
volume fraction of MX (V ) at TT; and (4) the volume fraction of M C precipitates (V M23C6 ) at TT.
MX
23 6