a National Key Laboratory of Science and Technology on Advanced Composites in Special Environments, Harbin Institute of Technology, Harbin 150001, China
Carbon/carbon-silicon carbide (C/C-SiC) composites are critical thermostructural materials for extreme aerospace environments, where flexural strength governs structural reliability. However, reliable prediction of flexural strength remains difficult because flexural strength is governed by complex nonlinear process-structure-property relationships, while experimental evaluation is time-consuming and costly. Traditional ML models often struggle to capture the complex structure-property relationships embedded in unstructured textual descriptions. Despite large language models (LLMs) excelling in natural language, their direct application to small-sample numerical regression remains challenging. This study introduces a semantics-guided multimodal machine learning (SGMML) framework using a LoRA-fine-tuned LLM text encoder with structured numerical features. The SGMML model based on a dataset of 142 cases achieved an MAE of 25.96 MPa and an R2 of 0.81 outperforming traditional ML models by 20.9% and fine-tuned LLMs by 92.9% in terms of R2. Experimental validation on two C/C-SiC composites manufactured via different processes yielded prediction residuals of 2.1 and 1.5 MPa. Furthermore, independent validation on external literature samples yielded an MAE of 18.08 MPa, demonstrating the framework's transferability beyond the training data. These results show that combining structured descriptors with domain-specific textual encoding enables accurate and scalable property prediction for advanced composites under data-limited conditions.
Composites are engineered materials made up of two or more different components, usually a matrix and a reinforcement, with the goal of achieving better properties than either component alone. Carbon/carbon-silicon carbide (C/C-SiC) composites, characterized by a carbon and silicon carbide dual-matrix, combine the high strength and low density of carbon fibers with the oxidation resistance and high-temperature stability of silicon carbide matrix [1], [2], [3]. They are widely regarded as promising thermostructural materials for aerospace and other extreme-environment applications because of their low density, high specific strength, and stable performance at elevated temperatures [4]. In practical service under extreme environments, these materials are often subjected to coupled thermal and mechanical loading, and their structural integrity is therefore closely related to component reliability [5], [6]. Among the relevant mechanical properties, flexural strength is of particular importance because it reflects the load-bearing capacity and damage tolerance of the composite under bending conditions that are frequently encountered in structural components [7], [8], [9].
Nevertheless, the reliable prediction of the flexural strength of C/C-SiC composites remains challenging because this property is governed by complex process-structure-property relationships spanning multiple length scales [10], [11]. The final mechanical performance depends not only on constituent composition and reinforcement architecture, but also on processing route, densification behavior, residual porosity, matrix distribution, and defect evolution [12], [13], [14], [15], [16]. Different fabrication routes, such as polymer infiltration and pyrolysis (PIP), chemical vapor infiltration/deposition (CVI/CVD), and reactive melt infiltration (RMI), can lead to significant differences in flexural behavior even for materials with similar nominal compositions [17], [18], [19]. Although experimental testing remains the most direct and reliable approach for evaluating flexural strength, it is generally costly, time-consuming, and labor-intensive, especially when broad process and parameter spaces must be explored. This encourages the application of data-driven approaches to speed up materials optimization and property prediction [20], [21], [22].
With the rapid development of materials informatics, machine learning (ML) has become an effective approach for establishing predictive relationships between input descriptors and target properties in materials science [23]. For structured tabular datasets, classical machine learning algorithms such as support vector regression (SVR), random forest (RF), and XGBoost have demonstrated strong capability in modeling nonlinear relationships with relatively low computational cost [24], [25]. These methods are particularly suitable for the small- and medium-sized datasets frequently encountered in materials research, where fully data-intensive models are often impractical [26], [27]. However, conventional machine learning models usually rely on manually defined numerical descriptors, which may limit their ability to represent more complex and context-dependent information [22]. In studies of materials processing, important information is often embedded in unstructured or semi-structured text, such as fabrication descriptions, process records, and microstructural observations [28], [29]. Such textual information may contain implicit knowledge related to defect formation, phase evolution, densification behavior, and expert interpretation, which is difficult to capture using a limited set of handcrafted numerical variables alone [30], [31], [32]. As a result, although conventional machine learning models can capture overall trends from structured data, they may still have limited ability to fully describe complex process-structure-property relationships when crucial semantic information is omitted.
Recent advances in large language models (LLMs) provide a potential route for incorporating textual information into materials property prediction [33], [34]. Compared with handcrafted numerical descriptors, text can convey richer contextual information regarding processing conditions, microstructural evolution, defect characteristics, and qualitative scientific interpretation [28], [29], [35]. Recent studies have suggested that language-based models can learn useful scientific representations when domain texts contain implicit links among terminology, processing concepts, and material behavior [33], [35], [36]. Still, the direct application of LLMs to numerical regression under the small-data conditions typical of advanced composites research remains insufficiently explored and validated, particularly in zero-shot and few-shot settings [26]. Parameter-efficient fine-tuning strategies, such as low-rank adaptation (LoRA), offer a practical way to adapt large pretrained language models to downstream tasks at relatively low cost [37]. Nevertheless, text-only LLMs may not necessarily replace the advantages of structured machine learning models in learning stable numerical relationships from limited experimental data [33]. Therefore, rather than viewing classical machine learning and LLMs as competing routes, a more promising strategy is to combine their complementary strengths within a multimodal framework [38], [39], [40]. To the best of our knowledge, systematic comparisons of classical machine learning, LLM adaptation, and semantic-structured fusion for predicting the flexural strength of C/C-SiC composites used in extreme environments remain limited.
Historical experimental data on C/C-SiC composites is underutilized. Extracting implicit process-structure-property knowledge from data-driven models allows for quick screening of manufacturing methods, considerably decreasing trial and error in materials creation. To address this issue, this work proposes a semantics-guided multimodal machine learning (SGMML) framework to predict the flexural strength of C/C-SiC composites, as illustrated in Fig. 1, in order to overcome the limits of purely numerical descriptors. The proposed approach employs a fine-tuned LoRA LLM to extract implicit microstructural semantics from manufacturing process texts, further integrating them with structured numerical aspects. Classical ML models are first benchmarked using structured numerical descriptors to evaluate their predictive performance and limitations. Subsequently, the regression capability of LLMs is assessed under zero-shot, few-shot, and LoRA-based fine-tuning settings to examine their applicability to low-sample numerical prediction in this task. Based on these results, a multimodal fusion framework is developed by combining textual semantic information extracted by an LLM with a traditional ML predictor. Finally, experimental validation and microstructural characterization are performed on C/C-SiC composites fabricated via PIP+CVD and PIP+RMI routes to verify the validity, robustness, and generalization capability of the proposed framework. The present study provides a practical route for integrating numerical descriptors and domain text semantics to model complex process-structure-property relationships in advanced ceramic-matrix composites.
2. Experimental procedures and methods
2.1. Data collection and preprocessing
The generalized C/C-SiC composite system studied in this work includes both bulk composites, where the matrix consists entirely of SiC and C, and surface-modified composites, in which SiC is deposited on the surface or shallow subsurface of porous C/C preforms by means of CVD/CVI to obtain localized densification. The “coating thickness” and textual “manufacturing process descriptions” were the essential input data used for the purpose of meticulous differentiation of the different mechanical behaviors and strengthening mechanisms of these two groups. This approach allows the language model to learn independently and distinguish between the global load-transfer processes of bulk matrices and the local strengthening effects of surface coatings. The dataset used in this study was constructed from the collected C/C-SiC composite records by combining structured process-related numerical descriptors with text-based material descriptions. Each sample contains quantitative features associated with fiber content, SiC volume fraction, density, porosity, coating thickness, and treatment temperature, together with encoded categorical descriptors representing fiber type, preform structure, manufacturing process, and defect category, as shown in Table 1.
The structured input vector had ten descriptors, including six continuous and four encoded category variables, while the goal variable was flexural strength. After preprocessing and unification, a total of 142 valid samples were retained for subsequent model development and evaluation. To ensure a fair comparison across all modeling strategies, including conventional machine learning, large language models, and the proposed SGMML framework, the same fixed train-test partition was adopted throughout this work.
For each sample, both category descriptors (e.g., manufacturing process) and quantitative features (e.g., fiber content and density) were extracted from the original literature. It should be noted that the “Defect type” (Dt) feature describes the composite before testing, thus avoiding any post-test data leakage. The defect descriptions were drawn from the original authors' pre-test characterizations, which often included non-destructive evaluation or microstructural observation of the as-fabricated specimens prior to any mechanical force being applied.
Rather than manual numerical categorization, the "Defect type" (Dt) feature uses textual descriptions derived straight from the microstructural characterization sections of the source material, keeping the authors' professional terminology. For a PIP-SiC modified C/SiC-C/SiC composite [41], the verbatim description "macro pores not fully filled with molten Si; SiC contains micro pores" was kept as the straight LLM input. The model uses this text-driven process to dynamically identify and weight complex microstructural properties, distinguishing phrases like "microcracks" and "delamination" from "dense" or "isolated pores. The missing numerical values were imputed with the mode number. The categorical variables were considered with label encoding. Given the small size of the dataset and the objective of developing a highly practical, domain-specific model, a fixed 80:20 hold-out split was employed. This yielded 113 training samples and 29 independent test samples, establishing an isolated test bed for final evaluation. Furthermore, 5-fold cross-validation was performed on the training set to ensure rigorous hyperparameter tuning (Fig. S1).
2.2. Experimental fabrication
Two example C/C-SiC composites with essentially distinct strengthening mechanisms were made and tested in order to verify the physical rationale of the suggested SGMML framework, as illustrated in Fig. 2. Sample A (PIP+CVD) represents a surface-strengthened system. A 2.5D preform (Vf=45\%, XY pitch-based / Z-direction T700 carbon fibers) was initially densified using pressure-assisted pitch infiltration and pyrolysis to reach a high density of 2.0 g/cm³ with minimal open porosity. A 20 μm SiC coating was subsequently produced using chemical vapor deposition (CVD) at 1200 °C for 12 hours, resulting in a final density of 2.1g/cm3 and a porosity of 5%. In contrast, Sample B (PIP+RMI) represents a bulk heterogeneous matrix system. A 2.5D preform (Vf=30\%, T700 fibers) was infused with sugar hydrogel pyrolyzed carbon (SC) to achieve a C/C composite with reduced intermediate density of 1.2g/cm3. The porous substrate was fully immersed in silicon powder and subjected to reactive melt infiltration (RMI) at 1650 °C for 2 hours. The liquid silicon entered the pores and interacted in situ with the carbon matrix, according to reaction (1) resulting in a final density of 2.3 g/cm3 and a porosity of 15%.
The flexural strengths of both composites were evaluated by carrying out three-point bending tests on a universal testing machine. The sample dimensions were 25 mm × 4 mm × 6 mm with a span of 20 mm and a crosshead speed of 0.5 mm/min. The phase compositions were analyzed by X-ray diffraction (XRD) and the fracture morphologies were observed by scanning electron microscopy (SEM). These two manufacturing methods were chosen to exemplify material systems commonly utilized in different extreme-environment conditions: surface coatings for oxidation and ablation protection, and bulk heterogeneous matrices for structural load-bearing at high temperatures.
2.3. Machine learning and large language models
2.3.1. Computational setup
After establishing a thorough dataset and synthesizing physical validation samples, a set of computational models was created to map the complex process-property relations. All experiments of this work were conducted on a workstation with the following configuration. Hardware configuration is 22 vCPU Intel(R) Xeon(R) Platinum 8470Q with one NVIDIA RTX PRO 6000 graphics card (1×96 GB VRAM) and 110 GiB of memory. And the software environment is shown in Table S1.
2.3.2. ML models
In this work, eight typical machine learning models [42], [43] including linear regression (LR), Ridge, Lasso, k-nearest neighbors (KNN), support vector regression (SVR), random forest (RF), gradient boosting regression (GBR), and XGBoost (XGB) were employed to predict the flexural strength of composites. Grid search and 5-fold cross-validation hyperparameter tuning were conducted for all models presented in Table S3 and Table S4 to get the ideal model configuration.
2.3.3. LLM models
In this study, several sample LLMs [41], [42] were used to predict the flexural strength using zero-shot, few-shot learning, and supervised fine-tuning. These models included both open-weight and closed-weight variants. We evaluated a range of models, and their details are summarized in Table 2. The models include Qwen-series models (0.8B, 2B, 4B, 9B, and Plus (API model ID: Qwen-plus)), DeepSeek-V3.2, ChatGPT-5.4, Claude-Opus-4.6, and Gemini-3.1 Pro Preview. All models were accessed or downloaded on April 20, 2026, following publicly available versions. For models without a formal technical report, official documentation, model cards, or release notes were used as a reference. Closed-source models (ChatGPT, Claude, and Gemini) were predicted using their official APIs, with the temperature parameter set to 0.0 to ensure deterministic results. Open-weight models (Qwen series) were tested using zero-shot, few-shot, and supervised fine-tuning conditions.
Zero-shot learning: predictions made without any composite samples.
Few-shot learning: prediction with few examples.
Supervised fine-tuning (SFT): The models are fine-tuned with the full training data set to the specific task of flexural strength prediction.
The prompt consists of a strict system instruction, optional few-shot demonstrations, and the target material features populated with sample-specific placeholders (bold bracketed terms). To conserve space, only key representative parameters are shown in Fig. 3. Full text prompt for both zero-shot and few-shot tasks are provided in Fig. S2.
2.4. Proposed SGMML Framework
The SGMML framework was proposed, which incorporates heterogeneous data sources through a three-layer architecture, as demonstrated in Fig. 4. This architecture includes textual encoding, feature compression and fusion, and XGBoost-based regression prediction.
Layer 1: Textual encoding. Unstructured process descriptions, such as fiber type, preform structure, fabrication route, and defect type, are tokenized and encoded by a pretrained language model named Qwen3.5-0.8B. In order to facilitate efficient domain adaptation in the presence of limited labeled data, Low-Rank Adaptation (LoRA) is implemented on the attention weight matrices. The adapted weight is as follows:
with $\mathit{B}\in {\mathit{R}}^{\mathit{d}\times \mathit{r}}$, $\mathit{A}\in {\mathit{R}}^{\mathit{r}\times \mathit{k}}$, and rank r=8, updating only ~1% of total parameters while keeping Wo frozen. The adaptation utilized a dropout rate of 0.10 to prevent overfitting. The hidden states from the final transformer layer are aggregated via mean pooling to obtain the text embedding ${\mathit{e}}_{\text{t}}\in {\mathit{R}}^{1024}$. In parallel, ten numerical features (e.g., fiber volume fraction, density, porosity) are standardized via Z-score normalization to yield ${\mathit{x}}_{\mathit{n}\mathit{u}\mathit{m}}\in {\mathit{R}}^{10}$.
Layer 2: Feature Compression and Fusion. The high-dimensional text embedding is first reduced by mutual information (MI) selection, retaining the top K=32 dimensions most relevant to the target property. The PCA model was only fitted on the training set in order to avoid any possible data leaking. The ideal number of dimensions was automatically determined using a 90% cumulative explained variance threshold, leading to dp=8 principal components. The fused feature vector is formed by early concatenation:
Layer 3: Regression Prediction. The fused feature vector z is directly input into an XGBoost regressor to forecast the flexural strength of C/C-SiC composites. XGBoost constructs an additive ensemble of regression trees, with the final prediction articulated as:
where zi is the fused feature vector of the i-th sample, T is the total number of regression trees, and ft denotes the t-th regression tree.
1. 2.5. Evaluation metrics
The predictive performance and generalizability of the models were quantitatively evaluated using five standard regression metrics [40]: coefficient of determination (R2), mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE). Mathematical definitions and discussion of these metrics can be found in Note S2 of the supplementary information.
3. Results and discussion
2. 3.1. Benchmark Evaluation: Limitations of Standalone Modalities
The standalone performance of structured-descriptor ML models and text-based LLMs was examined to set a baseline and justify the need for multimodal fusion. Flexural strength, as illustrated in Fig. S3, follows a right-skewed non-Gaussian distribution with a large tail toward high-strength samples. The mixed-type association analysis in Fig. S4 demonstrates that, while several process- and structure-related categorical descriptors, such as fiber type and manufacturing process, have relatively strong relationships with flexural strength, no single structured descriptor can fully explain its variation. These findings indicate that the flexural strength of C/C-SiC composites is driven by heterogeneous and nonlinear process-structure-property interactions, which supports the application of nonlinear modeling and multimodal fusion. As shown in the entire benchmark analysis, Fig. 5, which combines the performance indicators of both traditional ML models and LLMs, depending simply on structured numerical descriptors or pure text, limits the predictive boundaries. Table S6 and Fig. S5 describe the test performance of the eight standard machine learning models. Although tree-based ensemble models, such as XGBoost (R2=0.67), showed reasonable predictive capacity, they all underestimated high-strength samples (Fig. S6, Fig. S7). From a materials science standpoint, numerical descriptors like bulk density and nominal porosity only give macro-scale homogeneity of the composite state. They fail to discriminate between the various topologies of internal defects, such as whether a certain porosity appears as isolated micro-voids or interconnected micro-cracks. Such topological details are crucial for fracture initiation and propagation, influencing the ultimate flexural strength. This consistent bias suggests that handcrafted numerical descriptors alone are insufficient to reflect the intricate strengthening mechanisms and intrinsic defect evolutions found in high-performance outliers.
LLMs were investigated encoding process-related textual records in order to close this microstructural context gap. Fig. 5b shows that zero-shot and few-shot LLMs struggled with numerical regression, resulting in high variance and low R2 scores (Fig. S8). The predictions were substantially stabilized by LoRA; however, the standalone text modality still failed to meet the necessary accuracy for engineering applications. Despite their exceptional ability to comprehend semantic contexts, LLMs are limited in their precision in continuous numerical space by their inherent nature as autoregressive language models. Traditional machine learning models are numerically accurate but microstructurally blind, while stand-alone LLMs are semantically aware but numerically imprecise. This clearly shows that the structure-property mapping in small-sample composite datasets cannot be fully resolved by either solely numerical features or raw text production, which strongly motivates the new framework.
3.2. Performance and Interpretability of the Proposed SGMML Framework
The SGMML approach incorporates both modalities by utilizing the complementary benefits of text-derived structural awareness and robust numerical regression. Fig. 6 shows the SGMML model's parity diagrams. The anticipated values roughly approximate the ideal diagonal trend. The model performed well on the test set, with an R2 of 0.81, RMSE of 42.31 MPa, and MAE of 25.96 MPa. The severe underestimation in the high-strength region was substantially reduced in comparison to the baseline models.
An extensive ablation study was performed to validate the architectural decisions within the SGMML framework (Table 3). Comparison of variant A1 (numeric only) and A2 (text only) shows that text-only processing records have a higher predictive potential (R2=0.774) than numeric metrics only (R2=0.710). This is a surprising but important result in our dataset. However, naïve concatenation of raw text and numbers (Variant A3) leads to performance degradation, showing that cross-modal semantic interference takes place without the proper encoding of representations. The introduction of LoRA adaptation (Variant B3) was crucial to converting generalized LLM embeddings into task-specific microstructural representations. In addition, the feature dimensionality reduction strategy (Group C) mitigated the risk of overfitting under small-sample conditions; mutual information selection alone accounted for most of the gain. The joint use of mutual information filtering and PCA (Variant C5) reduced the 1024-dimensional textual embeddings to 8 latent axes, successfully removing textual noise at the expense of core processing semantics. This delicate balance between retaining multi-modal information while rigorously controlling the complexity of the feature space explains why XGBoost (Variant D6) was able to successfully map the final non-linear correlations without overfitting.
To further analyze the superiority of the SGMML framework over numerical baselines, the latent feature spaces were visualized by t-SNE (Fig. 7a-e). The numeric-only format shows considerable overlap of samples at different intensity levels. On the contrary, the SGMML framework generates a well-organized latent space, where samples with the same flexural strength are evidently clustered. Moreover, the feature importance analysis (Fig. 7f) shows that the extracted first principal component from the LLM (LLM-PC1) is the most important predictor. Correlation analysis (Fig. 7g) shows that LLM-PC1 has strong relationships with the fabrication method and fiber type, acting as a coherent semantic axis to capture process-induced structural properties hidden inside the textual descriptions. This shows that the LLM has successfully created a semantic axis that connects processing methods to microstructural flaws. Therefore, the LLM-derived semantic axis links processing routes to defect-related descriptors, which is consistent with the known structure-property relationships of C/C-SiC composites: processing conditions determine defect formation, which in turn affects mechanical performance.
3.3. Independent validation of the SGMML framework
To ensure that the SGMML framework accurately captured the underlying process-structure-property interactions, rather than merely overfitting numerical correlations, two 2.5D C/C-SiC samples produced through fundamentally different methods were experimentally validated. The processing and material parameters of these two composite samples are detailed in Table 4.
Microstructurally, the PIP+CVD composite (Figs. 8a, 8b), which has a reduced SiC content and is reinforced with mixed pitch/T700 fibers, shows a thick, continuous matrix with few defects. In contrast, the PIP+RMI sample (Fig. 8d, e), consisting of high-strength T700 fibers with a much higher SiC content, presents network-like cracks, delamination, and increased porosity resulting from the intense 1650 °C reactive melt infiltration process. At such elevated temperatures, thermal degradation and silicon vapor erosion inevitably cause damage to the carbon fibers. The chemical and structural differences distinctly influence their macro-mechanical responses (Figs. 8c, 8f). The PIP+RMI composite showed higher average flexural strength than the PIP+CVD one due to the homogeneous T700 fibers, higher final density, and strong SiC-rich matrix. However, its substantial internal cracks and localized structural weaknesses led to a more brittle fracture mode with sudden load drops. On the other hand, the PIP+CVD sample, although less strong due to the limitations imposed by the pitch-based fibers and the limited SiC content, presented better damage tolerance and a more progressive failure process. This is directly due to the dense, continuous matrix of carbonaceous material and the lack of major pre-existing structural fractures. These results, confirmed by processing dependent XRD phase analysis, indicate that the changes in mechanical properties are governed by the combined effects of fiber selection, defect states, and matrix integrity. These results provide experimental evidence supporting the physical rationality of the proposed SGMML framework.
Fig. 9a and Fig. S9 compares the prediction errors of several models for the two experimental validation samples. The SGMML framework produced the lowest absolute residuals (2.1 MPa for Sample A and 1.5 MPa for Sample B), successfully staying within the limits of experimental uncertainty. In contrast, typical ML models consistently produced larger errors, and stand-alone LLMs produced erratic forecasts with severe under- or overestimation. To systematically examine the SGMML framework's generalization potential outside of its training area, 8 completely new C/C-SiC examples from published literature were used. These samples, which encompass a wide range of fiber architectures and processing routes, were wholly concealed during the model development process. SGMML's prediction performance on external samples was robust, with an MAE of 18.08 MPa, an RMSE of 25.76 MPa, and an R² of 0.70 (Fig. 9b and Table S9). The external MAE is comparable to the internal test MAE of 25.96 MPa, indicating that the prediction error does not inflate appreciably beyond the training domain. The lower external R² of 0.70 versus 0.81 mainly arises from the narrower strength range of the external samples rather than a degradation in predictive accuracy.
4. Conclusions
This study proposes an SGMML framework combining an LLM text encoder with a ML model to predict the flexural strength of C/C-SiC composites for extreme-environment applications. The main findings are summarized as follows:
1. Traditional machine learning models exhibit systematic biases for high-strength data due to the lack of microstructural context in numerical descriptors, while standalone LLMs are insufficient for accurate numerical regression.
2.The SGMML framework combines numerical characteristics with LoRA-encoded process semantics, yielding high predictive accuracy with an R2 of 0.81 and an MAE of 25.96 MPa.
3. Experimental validation on two composites with different strengthening procedures resulted in small prediction errors (residual < 2.5 MPa), demonstrating the model's physical rationale and practical reliability.
The proposed framework demonstrates practical utility for the targeted domain. Future work will focus on enlarging the text-numerical dataset to cover a wider range of extreme-environment conditions and to improve the model’s generalizability across different composite systems.
Q.Chen, J.Huang, D.Li, Y.Ye, S. Bai. The thermomechanical coupling and multiscale correlation mechanism of Cf/C-SiC composites reinforced with chopped carbon fibers. Composites, Part B, 2026, 315: 113513. https://doi.org/10.1016/j.compositesb.2026.113513
[2]
Y.Chen, R.S.M.Almeida, S.Flauder, et al. In-situ X-ray tomography and acoustic emission monitoring of damage evolution in C/C-SiC composites fabricated by liquid silicon infiltration. Advanced Science, 2025, 12: e16200. https://doi.org/10.1002/advs.202516200
T.Jing, Y.Xin, L.Zhang, X.Sun, C.He, et al. Application of carbon fiber-reinforced ceramic composites in active thermal protection of advanced propulsion systems: a review. Chemical Record, 2023, 23: e202300022. https://doi.org/10.1002/tcr.202300022
[5]
J.Qu, W.Wang, Y.Shangguan, et al. Thermomechanical modelling and strength assessment of C/C-SiC composite brake disc. Applied Composite Materials, 2023, 30: 1547-1568. https://doi.org/10.1007/s10443-023-10136-6
[6]
P.Xu, H.Zhang, E.Jin, Z.Tan, P.Zhang, Y.Li, et al. Thermo-mechanical coupling failure mechanisms of reusable SiCf/SiC composites with PyC interphase. International Journal of Fatigue, 2024, 189: 108580. https://doi.org/10.1016/j.ijfatigue.2024.108580
[7]
Y.Wang, Z.Guan, X.Wang, F.Liu, T.Jiang, J. Xu. Mechanical properties and damage analysis of C/C-SiC curved beam under four-point bending: experimental and numerical investigation. Ceramics International, 2020, 46: 30008-30019. https://doi.org/10.1016/j.ceramint.2020.07.040
[8]
W.Lu, Z.Wang, Y.Chen, Z. Qu. Flexural property and failure mechanism of C/SiC composites investigated by high-temperature three-point bending and in situ X-ray computed tomography. Journal of the European Ceramic Society, 2026, 46: 117952. https://doi.org/10.1016/j.jeurceramsoc.2025.117952
[9]
Y.Shi, Y.Xiu, P.Wu, C.Wang, Y.Xiao, R.Jemmali, et al. Notch sensitivity of C/C-SiC composite evaluated by flexural tests. Journal of the European Ceramic Society, 2024, 44: 2780-2789. https://doi.org/10.1016/j.jeurceramsoc.2023.12.075
[10]
J.J.Sha, S.H.Wang, J.X.Dai, Y.F.Zu, W.Q.Li, R.Y. Sha. Improved microstructure and high temperature mechanical properties of C/C-SiC composites by introduction of ZrC nanoparticles. Ceramics International, 2020, 46: 10054-10062. https://doi.org/10.1016/j.ceramint.2019.12.034
[11]
J.Hao, J.Li, X.Shi, P.He, W.Shi, Y.Tan. Changes in microstructure and mechanical properties of the carbon fiber and their effects on C/SiC composites. Materials Characterization, 2022, 193: 112334. https://doi.org/10.1016/j.matchar.2022.112334
[12]
Z.Ye, Y.Wang, Q.Wen, Y.Liu, C.Liu, et al. Effects of density and heat treatment of C/C preforms on microstructure and mechanical properties of C/C-SiC composites. International Journal of Applied Ceramic Technology, 2023, 20: 3386-3398. https://doi.org/10.1111/ijac.14232
[13]
C.Liang, X.Gao, L.Fu, H.Mei, L.Cheng, L.Zhang. Pore evolution and mechanical response under locally varying density defects in ceramic matrix composites. Composites, Part B, 2024, 279: 111459. https://doi.org/10.1016/j.compositesb.2024.111459
[14]
J.M.Hausherr, L.Wagner, M.Gorywoda. Fabrication of artificial defects and their effect on the mechanical properties of C/C-SiC. Journal of the European Ceramic Society, 2020, 40: 2154-2162. https://doi.org/10.1016/j.jeurceramsoc.2019.10.012
[15]
R.Zhao, S.Pang, C.Hu, J.Li, B.Liang, S.Tang, et al. Fabrication of C/SiC composites by siliconizing carbon fiber reinforced nanoporous carbon matrix preforms and their properties. Journal of the European Ceramic Society, 2023, 43: 2070-2079. https://doi.org/10.1016/j.jeurceramsoc.2022.10.028
[16]
L.Xi, S.Li, K.Xue, X.Cui, B.Wang, Y.Li, et al. Internal damage evolution of C/SiC composites in air at 1650 °C studied by in situ synchrotron X-ray imaging. Composites, Part B, 2025, 288: 111878. https://doi.org/10.1016/j.compositesb.2024.111878
J.Zhang, Y.Zhang, Y.Wang, F.Wan, J.Li, D.Li, et al. Long-term oxidation performance of SiCf/SiC composites at 1200 °C in air atmosphere manufactured by PIP and hybrid CVI/PIP techniques. Ceramics International, 2024, 50: 13306-13315. https://doi.org/10.1016/j.ceramint.2023.12.337
[19]
D.Ma, Q.Liu, S.Ma, Y.Liu, J.Li, X.Xu, et al. Preparation of C/SiC composites by PIP and RMI joint processes: tunable matrix heterogeneity and toughening mechanisms. Journal of the European Ceramic Society, 2025, 45: 117665. https://doi.org/10.1016/j.jeurceramsoc.2025.117665
K.Choudhary, B.DeCost, C.Chen, A.Jain, et al. Recent advances and applications of deep learning methods in materials science. npj Computational Materials, 2022, 8: 59. https://doi.org/10.1038/s41524-022-00734-6
[22]
S.Kadulkar, Z.M.Sherman, V.Ganesan, et al. Machine learning-assisted design of material properties. Annual Review of Chemical and Biomolecular Engineering, 2022, 13: 337-360. https://doi.org/10.1146/annurev-chembioeng-092220-024340
[23]
C.Li, K.Zheng. Methods, progresses, and opportunities of materials informatics. InfoMat, 2023, 5: e12425. https://doi.org/10.1002/inf2.12425
[24]
G.Huang, Y.Guo, Y.Chen, Z.Nie. Application of machine learning in material synthesis and property prediction. Materials, 2023, 16: 5977. https://doi.org/10.3390/ma16175977
[25]
A.Baral, M.A.Chowdhury, F.A.Rafi, et al. A diverse and explainable machine learning framework for multi-property prediction in nanoparticle-enhanced carbon fiber/epoxy composites. Composites, Part B, 2026, 312: 113310. https://doi.org/10.1016/j.compositesb.2025.113310
S.K.Achar, J.A.Keith. Small data machine learning approaches in molecular and materials science. Chemical Reviews, 2024, 124: 12608-12671. https://doi.org/10.1021/acs.chemrev.4c00957
[28]
E.A.Olivetti, J.M.Cole, E.Kim, O.Kononova, G.Ceder, et al. Data-driven materials research enabled by natural language processing and information extraction. Applied Physics Reviews, 2020, 7: 041317. https://doi.org/10.1063/5.0021106
[29]
O.Kononova, T.He, H.Huo, A.Trewartha, E.A.Olivetti, et al. Opportunities and challenges of text mining in materials research. iScience, 2021, 24: 102155. https://doi.org/10.1016/j.isci.2021.102155
[30]
W.Wang, X.Jiang, S.Tian, P.Liu, D.Dang, Y.Su, et al. Automated pipeline for superalloy data by text mining. npj Computational Materials, 2022, 8: 14. https://doi.org/10.1038/s41524-021-00687-2
[31]
A.R.Durmaz, A.Thomas, L.Mishra, R.N.Murthy, T.Straub. An ontology-based text mining dataset for extraction of process-structure-property entities. Scientific Data, 2024, 11: 1277. https://doi.org/10.1038/s41597-024-03926-5
[32]
L.Montanelli, V.Venugopal, E.A.Olivetti, et al. High-throughput extraction of phase-property relationships from literature using natural language processing and large language models. Integrating Materials and Manufacturing Innovation, 2024, 13: 547-562. https://doi.org/10.1007/s40192-024-00344-8
[33]
G.Lei, R.Docherty, S.J.Cooper. Materials science in the era of large language models: a perspective. Digital Discovery, 2024, 3: 1708-1723. https://doi.org/10.1039/D4DD00074A
[34]
A. NiyongaboRubungo, C.Arnold, B.P. Rand, A.B.Dieng. LLM-Prop: predicting the properties of crystalline materials using large language models. npj Computational Materials, 2025, 11: 186. https://doi.org/10.1038/s41524-025-01536-2
[35]
X.Jiang, W.Wang, S.Tian, H.Wang, et al. Applications of natural language processing and large language models in materials discovery. npj Computational Materials, 2025, 11: 138. https://doi.org/10.1038/s41524-025-01554-0
[36]
T.Gupta, M.Zaki, N.M.A.Krishnan, Mausam. MatSciBERT: a materials domain language model for text mining and information extraction. npj Computational Materials, 2022, 8: 102. https://doi.org/10.1038/s41524-022-00784-w
[37]
J.E.Hu, Y.Shen, P.Wallis, Z.Allen-Zhu, Y.Li, S.Wang, et al. LoRA: low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021.
[38]
V.Moro, C.Loh, R.Dangovski, A.Ghorashi, A.Ma, Z.Chen, et al. Multimodal foundation models for material property prediction and discovery. Newton, 2025, 1: 100016. https://doi.org/10.1016/j.newton.2025.100016
[39]
Y.Wu, M.Ding, H.He, Q.Wu, S.Jiang, et al. A versatile multimodal learning framework bridging multiscale knowledge for material design. npj Computational Materials, 2025, 11: 267. https://doi.org/10.1038/s41524-025-01767-3
[40]
L.Song, D.Wang, X.Liu, A.Yin, Z.Long. Prediction of mechanical properties of composite materials using multimodal fusion learning. Sensors and Actuators A: Physical, 2023, 358: 114433. https://doi.org/10.1016/j.sna.2023.114433
[41]
X.Ma, H.Sun, S.Kou, S.Fan, J.Deng, L.Zhang, et al. Flexural strength and wear resistance of C/C-SiC brake materials improved by introducing SiC ceramics into carbon fiber bundles. Ceramics International, 2021, 47: 24130-24138. https://doi.org/10.1016/j.ceramint.2021.06.014
T.Li, Y.Zhang, Y.Fu, J.Sun, J.Li. Siliconization elimination for SiC coated C/C composites by a pyrolytic carbon coating and the consequent improvement of the mechanical property and oxidation resistances. Journal of the European Ceramic Society, 2021, 41: 5046-5055. https://doi.org/10.1016/j.jeurceramsoc.2021.04.008
F.Li, P.Xiao, Z.Li, J.Li, P.Deng, W.Liang, et al. Microstructures and properties of C/C-SiC composites derived from different gaseous carbon matrix precursors. Ceramics International, 2021, 47: 30777-30789. https://doi.org/10.1016/j.ceramint.2021.07.258
[47]
Y.Li, P.Xiao, Y.Shi, R.S.M.Almeida, W.Zhou, Z.Li, et al. Mechanical behavior of LSI based C/C-SiC composites subjected to flexural loadings. Composites, Part A, 2017, 95: 315-324. https://doi.org/10.1016/j.compositesa.2017.01.024
[48]
Z.Ye, Y.Wang, X.Xiong, J.Wang, T.Li, H.Liu, et al. Microstructure, interfacial and mechanical properties of SiC interphase modified C/C-SiC composites prepared by reactive melt infiltration. Journal of the European Ceramic Society, 2024, 44: 116785. https://doi.org/10.1016/j.jeurceramsoc.2024.116785
[49]
S.Flauder, N.Langhof, W.Krenkel, S.Schaffner. Size effect of carbon fiber-reinforced silicon carbide composites (C/C-SiC): Part 1 - bending load and statistical effects. Journal of the European Ceramic Society, 2021, 41: 6805-6816. https://doi.org/10.1016/j.jeurceramsoc.2021.07.040
[50]
W.Guo, S.Bai, Y. Ye. Controllable fabrication and mechanical properties of C/C-SiC composites based on an electromagnetic induction heating reactive melt infiltration. Journal of the European Ceramic Society, 2021, 41: 2347-2355. https://doi.org/10.1016/j.jeurceramsoc.2020.11.018
RIGHTS & PERMISSIONS
Just Accepted
This article has successfully passed peer review and final editorial review, and will soon enter typesetting, proofreading and other publishing processes. The currently displayed version is the accepted final manuscript. The officially published version will be updated with format, DOI and citation information upon launch. We recommend that you pay attention to subsequent journal notifications and preferentially cite the officially published version. Thank you for your support and cooperation.
PDF
