Prediction of thaw settling coefficient of frozen soil using machine learning techniques Authors: Ke Wang, Shuang Liu, Wenqiang Xing, LianZhen Zhang, Shuang Tian, Mengyao Qiu, Shanzhen Li Published: 2025 ABSTRACT AbstractThe thaw settling coefficient (TSC) of frozen soil serves as a crucial metric for assessing frozen soil thawing deformation. This work investigates the relationship of four parameters - freezing temperature (T), dry density (), water content (w), and overlying load (P) with the TSC of frozen soil. Utilizing a nonlinear prediction approach, multiple models are employed to predict the TSC values of frozen soil under varying influencing factors. Leveraging an experimental database comprising 841 data samples, the study conducts a series of sensitivity analyses to identify the most influential parameters in each nonlinear model and to determine the optimal performing model. Results from artificial neural network (ANN) and adaptive neuro-fuzzy inference system (ANFIS) models exhibit superior accuracy in predicting the TSC of frozen soil. The ANFIS model undergoes validation twice using new databases, demonstrating its reliability in meeting engineering requirements. Employing the mutual information (MI) analysis method, the study quantifies the impact of different influencing factors on the TSC of frozen soil, offering valuable insights for predicting TSC of frozen soil.Keywords: Thaw settling coefficient, Frozen soil, Predictive models, Machine learningSubject terms: Engineering, Environmental sciences Introduction The thaw-induced deformation of frozen soil, triggered by both climatic fluctuations and engineering activities, poses significant constraints on the advancement of high-speed railways, highways, and other infrastructure projects1–5. Notably, the Harbin to Yichun high-speed railway represents a pioneering endeavor in China, traversing a vast expanse of discontinuous island permafrost6. Achieving millimeter-level deformation control along high-speed railways necessitates the preservation of permafrost stability beneath the railway base7. However, the anticipated thawing-induced deformation of frozen soil in this locale poses formidable challenges to the operational viability of the high-speed railway. Given the extensive distribution of frozen soils and the rapid expansion of cold-region infrastructure in China, accurate and transferable prediction of thaw settlement parameters (e.g., TSC) is of direct importance for engineering design, risk control, and long-term maintenance. The thawing process in frozen soil entails the conversion of solid ice crystals into liquid water. During this transition, soil pores undergo compaction and discharge, consequently inducing vertical deformation. The characterization of the magnitude of deformation resulting from frozen soil thawing is encapsulated by the thaw settling coefficient (TSC) of frozen soil, as illustrated in Fig. 1. The calculation of the TSC of frozen soil is elucidated by Eq. (1). Definition of TSC and influencing factors on frozen soil. Previous research has demonstrated that pivotal determinants influencing the TSC of frozen soil encompass T, , w, and P8–18. Specifically, T and w delineate the extent of phase transition between water and ice within frozen soil, indicative of the unfrozen water content. serves as a proxy for the pore size distribution within frozen soil. The imposition of P signifies the external perturbation exerted during the consolidation of frozen soil. The TSC of frozen soil is typically measured through laboratory tests, which require significant time and effort. In particular, there are challenges associated with testing the TSC of undisturbed frozen soil, including low sampling rates and difficulties in maintaining a frozen state during transportation. Nevertheless, numerous tests have been conducted on the TSC of frozen soils across various soil types, yielding valuable insights18,19. However, a thorough analysis of preceding studies’ test results reveals discernible disparities in the TSC of frozen soil, even within the same soil type, owing to the discrete and stochastic nature of geotechnical particulate matter. Consequently, a more systematic investigation is imperative to ascertain the characteristics of thaw settlement of frozen soil. Previous research on the TSC of frozen soil predominantly falls into three categories: (a) laboratory experimentation, (b) empirical model fitting, and (c) machine learning. The first method employed to obtain the TSC of frozen soil involved conducting laboratory experiments to analyze the effects of T, , w, and P on the TSC of various types of frozen soil. Given China’s vast territory, variations in soil types contribute to differences in the TSC of frozen soil. Experimental tests were conducted on the TSC of frozen soil in various regions of China. In Northeast China, represented by the black soil of the Greater Khingan Mountains, these tests were carried out6,8,9,12,16. Similarly, investigations were conducted in Southern China, represented by Huainan soft soil10,11,13. Additionally, examinations were performed in Northwest China, characterized by loess in Shaanxi Province and silty clay in Inner Mongolia and Lanzhou14,15,17,20. Considerable time and sophisticated equipment were required for these tests, and the resulting data are presented in Fig. 2. The second approach to ascertain the TSC of frozen soil involves developing various empirical models based on laboratory test data. An extensive array of tests on the TSC of frozen soil was conducted by Chen8 and, based on the test results, a predictive formula for the TSC was proposed, considering variations in w and . Similarly, tests on the TSC of frozen soil were conducted by Wang6, and a predictive model for different types of frozen soil was devised. Tests on the TSC of frozen soil were conducted by Huang9, and a corresponding prediction formula was formulated based on the test data. These formulas elucidate the relationship between individual influencing factors and the TSC of frozen soil. The third approach for predicting the TSC of frozen soil involves machine learning (ML). This method can discern the intricate relationships among various factors affecting the TSC of frozen soil, as evidenced by a plethora of measured data, and utilize these relationships to accurately forecast the TSC. Consequently, the BP neural network algorithm was employed to predict the TSC of frozen soil, yielding insights that significantly enhance comprehension of frozen soil characteristics21. Moreover, there has been a notable surge in the development of algorithm models based on machine learning in recent years, a trend clearly observed within the domain of geotechnical engineering22–30. Recent machine-learning research in frozen-ground engineering has expanded to predicting unfrozen water content, forecasting permafrost-table or freeze-thaw states using climate and ground-temperature data, and mapping thaw settlement susceptibility/risk at regional scales31–36. Spatial distribution of thaw settling coefficient test data compiled from published studies. Most current studies predicting frozen soil TSC rely on empirical formulas, which suffer from low accuracy and limited practical applicability. Additionally, while some utilize artificial intelligence to construct prediction models, these often overlook soil type variations and lack mechanisms to analyze influencing factors. Given these limitations, this research primarily employs two distinct machine learning algorithms: Artificial Neural Network (ANN) and Adaptive Fuzzy Inference System (ANFIS). The model is trained, tested, and validated using 841 sets of the TSC of frozen soil test results. The study delves into the physical mechanisms underlying the influence of various parameters (T, , w and P) on the TSC of frozen soil. Through extensive data analysis, it determines the respective influence weights of each factor on the TSC, thereby enhancing researchers’ understanding of this critical permafrost behavior indicator. Notably, the experimental database compiled in this study synthesizes published laboratory tests on representative frozen soils across most major frozen-soil regions in China (Fig. 2), thereby providing direct engineering relevance for infrastructure design and maintenance in Chinese frozen-soil areas. Methods and material Background of ANN and ANFIS Artificial neural network (ANN) Artificial neural networks (ANNs) are data-driven models that can approximate complex nonlinear relationships without assuming an explicit functional form. In this study, ANN is used to learn the mapping between four commonly reported predictors (freezing temperature, dry density, water content, and overlying load) and the thaw settling coefficient (TSC) based on the compiled laboratory database. During training, the network parameters (weights and biases) are iteratively updated to minimize the discrepancy between measured and predicted TSC values. To improve robustness to occasional outliers in the compiled dataset, the Huber loss is adopted37. As illustrated in Fig. 3, each neuron computes a weighted sum of its inputs plus a bias, followed by a nonlinear activation function, which enables ANNs to capture multi-factor interactions. An artificial neuron j depicted with its various components. Further implementation details (data splitting, network configuration, training/testing/validation, and evaluation metrics) are provided in "Modelling process". Adaptive neuro-fuzzy inference system (ANFIS) Adaptive neuro-fuzzy inference systems (ANFIS) integrate fuzzy inference with neural learning to model nonlinear input-output relationships while retaining a rule-based structure. Following Jang’s formulation, ANFIS is adopted here as an alternative predictor of TSC, and its parameters are calibrated using the compiled dataset38. As shown in Fig. 4, ANFIS can be viewed as a five-layer architecture (fuzzification, rule firing, normalization, defuzzification, and output aggregation). In this study, membership functions and consequent parameters are learned from data, and the resulting model is evaluated using the same training–testing–validation procedure as ANN ("Modelling process"). Schematic representation of the five-layer architecture of ANFIS. Established database The most significant factor influencing permafrost TSC is the variation in soil composition. To establish a machine learning-based prediction model for permafrost TSC, it is imperative to incorporate input parameters that are readily attainable and applicable in engineering contexts. Given the above considerations, this study utilizes w, T, P, and as input parameters to predict the frozen soil TSC, effectively capturing the interplay of these key factors. To achieve the objectives of this study, data from various scholars’ research on frozen soils across different regions in China were referenced, as depicted in (Fig. 2). Leveraging the extensive expertise in the field and recognizing the diverse naming conventions present in previous literature concerning permafrost types, the dataset utilized comprises 841 sets of data categorized into clayey soil, silty soil, and sandy soil. This categorization was essential due to the diverse permafrost classifications encountered, which could vary up to 14 different types. Through a comprehensive review of experimental details from prior studies and leveraging years of experience in permafrost engineering, the author undertook the task of categorizing the dataset to streamline subsequent modeling efforts. Additionally, the clayey soil category encompasses clay and muddy soils, while the silty soil category includes silty clay and humus soil. Similarly, the sandy soil category comprises typical sandy soils and weathered sands. This comprehensive categorization enhances the precision and applicability of the dataset for subsequent modeling endeavors. Each dataset includes four input parameters for the input layer: w, T, P, and , alongside the TSC for the output layer. Given that the data utilized in model development stem from diverse ages, regions, and test conditions, it exhibits a certain degree of dispersion. Hence, prior to establishing the prediction model, preprocessing of the original data (data classification) is essential to ensure model accuracy and stability. This entails identifying and rectifying abnormal data within the database, which may arise from measurement fluctuations or testing inaccuracies. Moreover, for missing data from previous literature, the extent of missing data was determined through consultation with the authors of the previous studies and direct communication with them. The fundamental information of the selected database is summarized in Table 1, while additional details about the testing procedures can be found in the original research. Table 1Statistical information regarding variables used in this research to predict TSC.Statistical indexInput: water content, w (%)Input: temperature, T (℃)Input: loading, P (kPa)Input: dry density, (g/cm3)Output: thaw settlement coefficient, TSC, (%)Ave26.61– 14.750.61.485.24Standard deviation26.496.54208.710.277.56Min5.00– 35.00.000.210.0025% quantile19.39– 1501.41.350% quantile22– 1501.482.8175% quantile26– 1001.595.6Max330– 2.0010002.4858.0 Statistical information regarding variables used in this research to predict TSC. Rationale for input selection: This study focuses on freezing temperature (T), water content (w), dry density (ρd), and overlying load (P) because (i) these variables are consistently reported across most published TSC laboratory tests, enabling a unified compiled database; and (ii) they represent the primary physical controls on thaw settlement through phase change (T and w, linked to unfrozen water content), pore structure/porosity (ρd), and stress-driven thaw consolidation (P). Other potentially relevant variables (e.g., soil gradation, ice content, organic content, salinity, and freezing/thawing history) were not included here due to incomplete reporting in the source studies and are recommended for inclusion when data permit in future work. To provide a comprehensive overview of the dataset utilized in this machine learning study, histograms illustrating the distribution of all input parameters are presented in Fig. 5. Specifically, Fig. 5a to d depict the distribution of four input parameters for clayey soil, Fig. 5e to h for silty soil, and Fig. 5i to l for sandy soil. Histograms illustrating various variables utilized in this study. As depicted in Fig. 5, the distribution ranges for w (5% to 330%), T (-2℃ to 35℃), P (0 kPa to 1000 kPa), and dry density (0.21 g/cm3 to 2.48 g/cm3) among the four input variables exhibit substantial variability and non-normality. This wide-ranging distribution underscores the comprehensive and representative nature of prior research concerning the TSC of frozen soil. Furthermore, defining appropriate variable ranges for each input parameter is typically advisable to facilitate effective training of the machine learning (ML) model and enhance prediction accuracy. Indeed, the uneven distribution of these input parameters may yield more accurate predictions of TSC values compared to employing a highly restricted range of input parameters. To comprehensively analyze the fundamental attributes of the data used in machine learning, violin plots are employed to visually represent the variation range of four key test variables influencing the permafrost thaw and subsidence coefficient. Combining aspects of both box plots and kernel density plots, these visualizations offer a holistic view of the data distribution39. Within the violin plot, the median is indicated by a red dot, while the interquartile range is depicted by a gray bar, effectively illustrating the complete data distribution, as shown in Fig. 6. Violin plot for each parameter used in this research. In the data analysis process, assessing variable correlation is pivotal for gauging dataset integrity and establishing a robust framework for model development. To achieve this, correlation analysis was systematically executed on both input and output layer data, with specific attention given to individual soil categories, as depicted in Fig. 7. Heatmap of variable correlation. (a) Clayey soil, (b) Silty soil, (c) Sandy soil. The variable correlation falls within an acceptable range of -0.8 to 0.8, with analysis results indicating no discernible correlation among the three soil types indices categorized in this study. Based on these findings, each variable’s correlation is deemed satisfactory. Moreover, it is evident that the internal physical mechanisms underlying the influence of the four indices on the thawing coefficient of frozen soil vary significantly. Specific details regarding these differences in physical mechanisms will be elaborated upon in the subsequent sensitivity analysis section. Evaluation indices To evaluate the precision and reliability of the prediction model, four evaluation metrics were considered: Variance Account For (VAF), mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). The VAF, expressed as a percentage, measures how well the prediction aligns with the actual values by comparing the standard deviation of the fitting error to the standard deviation of the actual values. R² represents the square of the correlation coefficient between the predicted and observed values. RMSE indicates the standard deviation of the fitting error between the predicted and observed values, while MAE signifies the average discrepancy between them. Equation (2) to Eq. (5) are the calculation formulas for each evaluation metric. Specifically, the normalized RMSE (NRMSE, %) and normalized MAE (NMAE, %) are defined. Modelling process This section outlines the methodologies and modeling procedures employed to predict TSC values, utilizing two predictive techniques: Artificial Neural Networks (ANN) and Adaptive Neuro-Fuzzy Inference Systems (ANFIS). The discussion in this section will focus on the modeling process of these non-linear approaches, with the evaluation of their accuracy in predicting TSC to be addressed subsequently. ANN The initial step in modeling involves randomly partitioning the entire database into training, testing, and validation sets. The training dataset is utilized to train the model on various data patterns and learning processes, while the testing dataset evaluates the accuracy of the trained model. A smaller subset of the data is allocated for validation to ensure the model’s generalization and optimization. Following prior research practices, 70% (588 groups), 20% (168 groups), and 10% (85 groups) of the total dataset were designated as the training, testing, and validation sets, respectively. It is essential to highlight that the selection of samples for each set is randomized, ensuring an equal probability for each sample to be chosen. When employing ANN modeling, it is imperative to initially normalize the entire database within a constrained range (0–1). The selection of the number of hidden layers and nodes within each hidden layer is regarded as crucial in achieving an accurate MLP predictive model. As demonstrated by Hornik et al.40. A sufficient number of hidden nodes allows for the approximation of complex fitting problems by a single hidden layer in MLP modeling. For each MLP network, a single hidden layer with a sigmoidal activation function is utilized, accompanied by an output layer dedicated to TSC prediction. In prior investigations, the determination of the number of hidden nodes was often based on the count of input variables41,42. The upper limit of the suggested number of nodes is typically 2Ⅰ+1 (where Ⅰ represents the number of input variables) according to previous work43. Given the existence of 4 input variables in this study, the range considered for the number of nodes varies from 1 to 9. Subsequently, nine MLP models were constructed and their performance was assessed using R2 values, as illustrated in Table 2. The impact of the number of hidden nodes on the predictive performance of the model is evident. However, the model’s accuracy does not exhibit a continuous increase with the augmentation of node count. Selecting the most precise MLP model may present challenges due to the closely aligned R2 values. To address this issue, the approach proposed by Zorlu44 is adopted, which evaluates the collective performance of the model across training, testing, and validation phases. A hierarchical ranking based on R2 values is established, resulting in 9 MLP models being assigned corresponding levels. The highest R2 value (e.g., 0.981 for the training set) is assigned the highest level (level 9), while lower R2 values are assigned lower levels. In cases where different MLP models yield identical R2 values, they are assigned the same rank. The cumulative ranking of the training, testing, and validation sections for each MLP model ultimately determines the overall ranking. It is important to note that a higher score in the training set indicates stronger learning capability, while enhanced performance in the test and validation sets signifies greater generalization ability, rendering the model applicable in practical scenarios. In this study, hyperparameters were tuned using a hold-out training/testing/validation split and a ranking-based selection procedure44, rather than k-fold cross-validation, to preserve an independent test of generalization. We acknowledge that k-fold cross-validation can further reduce selection variance and will consider it in future work. Table 2Nine MLP models for predicting TSC of frozen soil.Hidden node number R 2 RankingTotal rankingTrainingTestingValidationTrainingTestingValidation10.6250.7050.815221520.6950.7160.848332830.7980.8950.8654431140.8590.9670.8955941850.9810.9580.9859892660.9810.9670.9019952370.9280.9480.9116661880.9480.9380.9757582090.9720.9540.97387722 Nine MLP models for predicting TSC of frozen soil. From the perspective of the total ranking value, model 5 has the best prediction performance when predicting TSC, and its ranking value accumulates to 26. The training, testing and verification of the MLP model established with 5 nodes are 0.981, 0.958 and 0.985, respectively. The above results show that the MLP model can predict the size of TSC value through four influencing factors. Note that the normalized measured and predicted TSC values have been normalized again to calculate additional performance prediction indices. The results and capabilities of artificial neural network models will be further discussed in a subsequent section. ANFIS ANFIS undergoes three essential stages: learning, testing, and validation. The selection of membership functions (MFs) holds particular significance in predicting TSC values using ANFIS. Previous studies have demonstrated the effectiveness of using Gaussian or trigonometric functions as membership functions due to their flexibility and simplicity45,46. Moreover, the number of membership functions significantly influences the model’s performance. In this study, six ANFIS models are constructed by employing two different sets of membership functions. With an increase in the number of MFs, the number of fuzzy rules also escalates. For instance, considering the four input variables in this study, the total number of fuzzy rules for MF = 3 and MF = 5 is 3^4 = 81 and 5^4 = 625, respectively. Therefore, to maintain a reasonable number of fuzzy rules, the authors utilize 3, 4, and 5 as input values for the number of MFs. The results exhibit similarity (Table 3), prompting the analysis of model performance based on R2 and RMSE values, mirroring the approach employed in the ANN section. The findings indicate that the ANFIS model featuring Trimf as the MF type and 5 MFs for each input variable is identified as the optimal predictive model for TSC. Hyperparameters of ANFIS (membership-function type and number of membership functions per input) were selected by comparing candidate configurations using the same training/testing/validation split and the ranking strategy, instead of k-fold cross-validation, to maintain an independent evaluation of generalization performance. Table 3Performance evaluation of ANFIS models with various specifications.ANFISMF typeNo.of MF in each inputNetwork performanceRankingTotal rankingR2R2R2RMSERMSERMSER2R2R2RMSERMSERMSETrainTestValidationTrainTestValidationTrainTestValidationTrainTestValidation1Gaussianmf30.8460.9040.9414.5123.4023.123111546182Gaussianmf40.8640.9250.9514.2323.7912.852235452213Gaussianmf50.8880.9460.9434.7202.8942.735453621214Trimf30.8730.9440.9434.1903.1292.916453333215Trimf40.9060.9880.9643.9072.8962.972566224256Trimf50.9150.9270.9463.7154.1153.12963416626 Performance evaluation of ANFIS models with various specifications. The range of input parameters utilized during the training phase is segmented into five categories: very low (VL), low (L), medium (M), high (H), and very high (VH). Furthermore, Table 4 displays some If-Then rules generated by the system during TSC prediction. The ANFIS model is trained to forecast TSC employing linguistic variables and If-Then rules, and the trained model is subsequently tested and validated using the same variables and rules. Table 4Examples of If-Then ANFIS rules for predicting TSC.If ( is VL) and (T is VL) and (w is VH) and (P is VH)Then (TSC is VH)If ( is L) and (T is L) and (w is H) and (P is VH)Then (TSC is H)If ( is M) and (T is M) and (w is M) and (P is M)Then (TSC is M)If ( is H) and (T is VH) and (w is VL) and (P is H)Then (TSC is L)If ( is VH) and (T is H) and (w is L) and (P is VL)Then (TSC is VL) Examples of If-Then ANFIS rules for predicting TSC. Results and discussion Model assessment A robust predictive model should demonstrate strong performance across training, testing, and validation stages. Table 5 presents the performance metrics considered during the training, testing, and validation of the nonlinear prediction model established in this study. The outcomes of each error index were consistent. While ANN exhibited favorable performance during the training phase, the results of the ANFIS model in the testing and validation stages closely aligned with the predicted TSC values. Table 5TSC results across training, testing, and validation phases.Evaluation indexML modelANNANFISTrainingTestingValidationTrainingTestingValidationR20.8940.9200.8690.8930.9330.881VAF (%)0.8970.9210.8700.8990.9330.888RMSE4.0754.0014.6063.9893.6064.286MAE3.0633.2233.6083.0692.9423.326NRMSE (%)7.0266.8987.9416.8786.2177.39NMAE (%)5.2815.5576.2215.2915.0725.734 TSC results across training, testing, and validation phases. To provide a comprehensive visual representation of the model prediction outcomes, the results from the training, testing, and validation phases of both the ANN and ANFIS models for the clay, silty soil, and sandy soil types are illustrated in Figs. 8, 9 and 10. These figures substantiate that both ANFIS and ANN models yield TSC predictions closely aligned with the measured values. Measured vs. predicted TSC for clayey soil: (a) to (c) focusing on ANN model, (d) to (f) focusing on ANFIS model. Measured vs. predicted TSC for silty soil: (a) to (c) focusing on ANN model, (d) to (f) focusing on ANFIS model. Measured vs. predicted TSC for sandy soil: (a) to (c) focusing on ANN model, (d) to (f) focusing on ANFIS model. To assess the predictive capabilities of the ANN and ANFIS models in estimating TSC, two additional prominent ML models, namely the Group Method of Data Handling (GMDH) and the Extreme Learning Machine (ELM), were also employed using the same training, testing, and validation datasets. While the construction of these models was based on their efficacy and influential factors, specific modeling details are omitted here as they are not the primary focus of this study. Taylor diagrams depicting the training, testing, and validation outcomes of the ANN, ANFIS, GMDH, and ELM models are presented in Figs. 11, 12 and 13, corresponding to clayey soil, silty soil, and sandy soil types, respectively. Taylor diagram for predictive models on clayey soil across all phases. Taylor diagram for predictive models on silty soil across all phases. Taylor diagram for predictive models on sandy soil across all phases. The Taylor diagrams provide an intuitive means to assess the reliability of the four models’ TSC prediction results. In these diagrams, closer clustering of data points towards the lower right quadrant indicates better predictive performance. Based on this criterion, ANFIS models demonstrate strong performance during the training phase, while ANN exhibits higher accuracy during testing and validation compared to the other models. Notably, both ANN and ANFIS models outperform the GMDH and ELM models. Further validation of the ANFIS model accuracy will be conducted, with new test data being incorporated. To further elaborate, the consistently higher proximity of ANN and ANFIS to the reference point in the Taylor diagrams confirms that these two models better balance correlation and error reduction than GMDH and ELM across different soil groups. In particular, ANN tends to provide strong fitting capability during training due to its flexible parameterization, whereas ANFIS offers slightly more stable generalization in several cases by embedding fuzzy rules that regularize the input–output mapping. From an engineering-use perspective, ANN can be preferred when sufficient representative data are available for the target site, while ANFIS is recommended when interpretability and robustness across heterogeneous datasets are prioritized. Validation with the other studies To further validate the model’s generalization capability, the authors conducted a secondary validation using permafrost data collected along the Harbin to Yichun high-speed railway. Utilizing the ANFIS model trained on clay, silt, and sandy soil, 10, 15, and 12 randomly selected data groups were respectively chosen for verification. The validation outcomes are depicted in Fig. 14. Predicted TSC values by the ANFIS model vs. measured TSC for secondary validation: (a) clayey soil, (b) silty soil, (c) sandy soil. The ANFIS model exhibits remarkable performance in predicting the TSC among three distinct soil types. Specifically, the predictive accuracy for sandy soil yielded the highest coefficient of determination (R2 = 0.99), followed by cohesive soil (R2 = 0.94) and silty soil (R2 = 0.92). Notably, ANFIS represents a fusion of Artificial Neural Networks (ANN) and fuzzy logic controllers, wherein the fuzzy rules are derived from ANN. This integration constitutes a fundamental distinction between ANN and ANFIS methodologies. While ANN’s controller blocks and parameters are algorithmically generated, ANFIS leverages a hybrid approach, combining the strengths of both ANN and fuzzy logic controllers. By incorporating neural networks into fuzzy systems, ANFIS not only augments predictive performance but also affords a more nuanced representation of internal information, owing to the learning capabilities inherent in neural networks47. Comparison of predicted TSC with previous models Empirical formulas derived from experimental data have been employed to estimate TSC6. This section contrasts the predictions generated by ANFIS with those obtained through two empirical formulas documented in prior literature. Equation (6) and (7) delineate the associations between TSC and w, as well as between TSC and , respectively. where a, b, c, and d represent the fitting parameters, each corresponding to specific coefficients, which can be acquired through reference sources. Figure 15 illustrates the comparison between the predicted results generated by ANFIS and the calculated outcomes obtained from Eqs. 15 and 16 across the test results of three soil types. The evaluation parameters of the models are presented in the figure. The results indicate that the TSC predicted by the ANFIS model closely aligns with the measured values compared to other models, affirming ANFIS as a more reliable model capable of estimating TSC values with higher accuracy. Comparison of measured and predicted TSC values using ANFIS and empirical equations. Sensitivity analysis To assess the significance of the input variables (P, T, and w) on TSC, the mutual information (MI) method was employed to determine the importance and sensitivity of each variable with respect to the TSC values. The MI method, serving as a filtering technique, can capture arbitrary relationships (both linear and nonlinear) between independent and dependent variables, thus providing an estimated measure of mutual information between each independent variable and the dependent variable48. It should be noted that the estimated value lies within the range [0, 1], where a value of 0 signifies an absence of correlation between the two variables, while a value of 1 indicates a strong positive correlation between them. The closer the estimated amount of an independent variable is to 1, the stronger its correlation with the dependent variable, and vice versa. The significance of these four input variables and TSC is illustrated in Fig. 16. Contribution of input parameters to TSC variation. Intuitively, among the input variables, w exhibited the strongest correlation with TSC, with a notable correlation index of 0.9437. The strong correlation observed underscores the significant influence of w on TSC values, consistent with the fundamental principles governing the thawing process. Elevated water content in the soil implies the presence of abundant ice crystals during freezing. These ice crystals exhibit greater volume compared to liquid water. Consequently, upon transitioning from ice to liquid state, rapid soil subsidence occurs. This phenomenon is intricately linked to the physical mechanisms governing frozen soil thawing. Following w, the and T displayed moderate correlations with TSC, with correlation indices of 0.3684 and 0.2909, respectively. These findings imply that both dry density and overlying load contribute to the variations observed in TSC, albeit to a lesser extent compared to water content. However, it is noteworthy that the correlation between P and TSC was found to be negligible, as evidenced by the low correlation index of 0.0011. This suggests that freezing temperature has a minimal influence on the TSC in the context of this study. Furthermore, it’s imperative to acknowledge a potential limitation in the analysis regarding the variable P. The conclusion regarding its minimal correlation with TSC may not be definitive, as the majority of the data used in model construction might have featured a consistent value of 0 for P. This aspect necessitates further investigation to ascertain the precise impact of overlying load on TSC across a broader range of conditions. Verification of this inference requires subsequent experimental results from TSC testing under varying overlying load conditions. Limitations and future works This study has several limitations that should be acknowledged. First, the compiled database integrates laboratory results reported by multiple studies, and thus unavoidable inter-study variability may exist due to differences in test protocols, instrumentation, and sample preparation. Second, although the dataset covers major frozen-soil regions in China and includes 841 laboratory data groups, the conclusions and model generalization are primarily supported within the Chinese context; extrapolation to other climatic or geological settings should be performed with caution. Third, the current framework considers four key predictors (T, ρd, w, and P) that are commonly available in the literature; however, additional variables that are physically relevant to thaw settlement may further improve prediction accuracy when such information is accessible. Despite these limitations, the proposed framework—trained and validated on a synthesized dataset representing major frozen-soil regions in China—provides a practical and efficient tool for rapid estimation of the thaw settling coefficient (TSC) in cold-region engineering applications. Future work may further enhance both predictive performance and practical applicability in the following directions: (i) expanding the database to include additional regions and adopting standardized laboratory protocols to reduce inter-study variability; (ii) incorporating more physically informative features, when available, such as soil gradation, ice content, organic content, salinity, and freezing/thawing history; (iii) adopting k-fold cross-validation for hyperparameter tuning and model selection to improve robustness and reproducibility; and (iv) introducing uncertainty quantification (e.g., prediction intervals or probabilistic outputs) to better support risk-informed engineering decision-making. In addition, comparative studies involving other machine-learning approaches (e.g., random forests and ensemble learning) are encouraged to benchmark accuracy and interpretability for thaw-settlement prediction. Conclusions This study develops data-driven models to predict the thaw settling coefficient (TSC) of frozen soil using four commonly reported predictors (freezing temperature, dry density, water content, and overlying load). Two machine-learning (ML) approaches—Artificial Neural Networks (ANN) and Adaptive Neuro-Fuzzy Inference Systems (ANFIS)—were trained, tested, and validated on a compiled database. The main conclusions are as follows: Both the ANN and ANFIS models demonstrate superior accuracy in predicting TSC compared to alternative ML models such as LEM and GMDH. Compared with empirical equations, the ANFIS-based predictions show systematically closer agreement with measured TSC values over a wide range of water content and dry density, highlighting improved transferability beyond single-factor fitting. Both the ANN and ANFIS models demonstrate superior accuracy in predicting TSC compared to alternative ML models such as LEM and GMDH. Compared with empirical equations, the ANFIS-based predictions show systematically closer agreement with measured TSC values over a wide range of water content and dry density, highlighting improved transferability beyond single-factor fitting. The ANN and ANFIS models exhibit robust performance across training and testing phases. Upon validation with new data, the ANFIS model consistently demonstrates reliable predictions, meeting engineering standards. Consequently, the ANFIS model emerges as a valuable tool for TSC prediction in frozen soil scenarios. The MI results also provide a data-driven basis for feature prioritization in future modeling, suggesting that water content should be treated as the primary driver, while the influence of overlying load requires additional targeted tests because of limited non-zero P samples in the current database. The ANN and ANFIS models exhibit robust performance across training and testing phases. Upon validation with new data, the ANFIS model consistently demonstrates reliable predictions, meeting engineering standards. Consequently, the ANFIS model emerges as a valuable tool for TSC prediction in frozen soil scenarios. The MI results also provide a data-driven basis for feature prioritization in future modeling, suggesting that water content should be treated as the primary driver, while the influence of overlying load requires additional targeted tests because of limited non-zero P samples in the current database. Validation against empirical formulas further confirms the efficacy of the ANFIS model in generating TSC estimates closely aligned with empirical observations. This validation underscores the potential wider adoption of ANFIS models among researchers and geotechnical engineers, subject to congruent conditions and meticulous consideration of input parameters and their ranges. Validation against empirical formulas further confirms the efficacy of the ANFIS model in generating TSC estimates closely aligned with empirical observations. This validation underscores the potential wider adoption of ANFIS models among researchers and geotechnical engineers, subject to congruent conditions and meticulous consideration of input parameters and their ranges. The Mutual Information (MI) analysis method is employed to ascertain the relative significance and sensitivity of various factors influencing the TSC of frozen soil. Notably, water content, or unfrozen water content, emerges as the predominant factor affecting TSC, followed by dry density and freezing temperature. However, additional data are warranted to conclusively evaluate the influence of overlying loads. The Mutual Information (MI) analysis method is employed to ascertain the relative significance and sensitivity of various factors influencing the TSC of frozen soil. Notably, water content, or unfrozen water content, emerges as the predominant factor affecting TSC, followed by dry density and freezing temperature. However, additional data are warranted to conclusively evaluate the influence of overlying loads. Future work should focus on enlarging and internationalizing the database, integrating additional influencing variables, and adopting cross-validation and uncertainty quantification to further enhance model robustness and facilitate engineering-grade application of ML-based TSC prediction. Future work should focus on enlarging and internationalizing the database, integrating additional influencing variables, and adopting cross-validation and uncertainty quantification to further enhance model robustness and facilitate engineering-grade application of ML-based TSC prediction. Acknowledgements This study was financially supported by the National Natural Science Foundation of China (Grant No. 42301138), National Natural Science Foundation of China (No. 42102311), the Fundamental Research Funds for the Central Universities (LH2022D016),and the Major Science and Technology Program of the Tibet Autonomous Region (XZ202402ZD0008). Author contributions KW participated in the writing of the manuscript, SL participated in the drawing of all the images, WQX, ST, and MYQ participated in the organization and analysis of the data, and LZZ and SZL provided guidance on the methods involved. Data availability Data will be made available on request. If needed, please contact kwang@szu.edu.cn. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement Data will be made available on request. If needed, please contact kwang@szu.edu.cn.