Method and system for developing large-model energizing heat-conducting asphalt material
By constructing a five-dimensional unified vector representation and large-scale model pre-training, combined with multi-task prediction and closed-loop iteration, the problems of long development cycle and high cost of traditional thermally conductive asphalt materials are solved. Efficient and accurate multi-objective optimization and risk warning are achieved, improving the R&D efficiency and interpretability of thermally conductive asphalt materials.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-30
- Publication Date
- 2026-04-03
AI Technical Summary
The development of traditional thermally conductive asphalt materials relies on empirical trial and error, which results in long R&D cycles, high costs, difficulty in handling multivariate coupling problems, and inconsistent data due to differences in equipment and testing conditions in different laboratories. Furthermore, the lack of a systematic framework makes it impossible to provide confidence levels and risk warnings, thus posing potential risks to engineering applications.
A five-dimensional unified vector representation containing multimodal data is constructed. Combined with multi-mechanism pre-training and multi-task prediction of a large thermally conductive asphalt model, accurate reverse design and closed-loop iteration are achieved. Through multi-task prediction, dynamic manufacturability constraints and closed-loop iteration mechanism, interpretable information is output to clarify key influencing factors and risk points.
It significantly shortens the R&D cycle, reduces costs, achieves precise optimization of multiple objectives, improves the accuracy of performance prediction, balances thermal conductivity, workability and carbon emission requirements, clarifies key influencing factors and risk points, and solves the problem of low efficiency in traditional experience-based trial and error.
Smart Images
Figure CN121789866A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent development technology for road engineering materials, specifically a method and system for developing thermally conductive asphalt materials using a large-scale model. Background Technology
[0002] Thermally conductive asphalt materials can effectively reduce peak pavement temperature, improve structural heat dissipation efficiency, and reduce thermal aging and rutting problems. They are widely used in bridge deck paving, tunnel pavement, airport runway, and roads in high-temperature areas. This material involves multiple components and various process variables, and needs to balance multiple objectives such as thermal conductivity, rheological properties, workability, and storage stability. There are complex coupling relationships between its multi-scale structure and multi-objective performance, making it a key research and development direction in the field of road engineering materials.
[0003] The development of traditional thermally conductive asphalt materials relies on empirical trial and error, requiring lengthy testing to verify effectiveness. This results in long development cycles, high costs, and difficulty in handling multivariate coupling problems. Differences in equipment and testing conditions between different laboratories lead to inconsistent data distribution, and traditional models have poor adaptability. Furthermore, multi-objective optimization lacks a systematic framework, and improvements in thermal conductivity are often accompanied by negative issues. The recommended formulation mechanisms are not clearly defined, and confidence levels and risk warnings cannot be provided, posing potential risks to engineering applications. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for the development of thermally conductive asphalt materials using a large model. By constructing a five-dimensional unified vector representation containing multimodal data, and combining multi-mechanism pre-training and multi-task prediction of the thermally conductive asphalt large model, it achieves accurate reverse design and closed-loop iteration, significantly shortens the R&D cycle, reduces costs, and outputs interpretable information to clarify key influencing factors and risk points.
[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On one hand, a method for developing large-scale thermally conductive asphalt materials, the specific steps of which are as follows:
[0006] Data processing: Construct a database containing formulation process data and multimodal characterization data, and encode the data into a five-dimensional unified formulation structure performance representation vector constrained by component process spectrum image. The vector contains a heat conduction network feature token obtained through the heat conduction network feature token encoding formula.
[0007] Model pre-training: Construct a large-scale basic model of thermally conductive asphalt containing an encoder and a task head, and train it using one or more mechanisms among mask reconstruction, contrastive learning, thermal conductivity and rheological property coupling constraint pre-training, and weakly supervised distillation. The coupling constraint pre-training is achieved through the core coupling constraint loss function.
[0008] Multi-task prediction: A multi-task prediction head is established based on a large model to output key performance indicators such as thermal conductivity, rheological properties, workability, storage stability, and aging sensitivity. One or more of the following strategies are used: ensemble model, Bayesian approximation, and Monte Carlo strategy, to output prediction uncertainty.
[0009] Reverse design and multi-objective optimization: Combining a four-dimensional dynamic manufacturability constraint library of raw materials, equipment costs, carbon emissions, etc., reverse design of formulation and process is performed through one or more of the following methods: Bayesian optimization based on surrogate model, Pareto front search based on evolutionary algorithm, and sequential decision based on constraint satisfaction. The output is K candidate formulations and process parameters.
[0010] Closed-loop iteration: The selection priority is calculated based on the three-dimensional test point selection priority formula. After the sample preparation and testing are completed, the data is fed back and the model parameters are adjusted through an incremental update mechanism to achieve closed-loop iteration.
[0011] Interpretable output: Output at least one interpretable piece of information from the following: ranking of the contribution of formulation variables, structural evidence, contribution indications of key spectral peak curve segments, and anomaly detection results.
[0012] Furthermore, in the data processing, the five-dimensional unified formulation structure performance representation vector includes: component token, process token, spectral curve token, image token, and constraint token. The calculation formula for the heat conduction network feature token is: ,in, This is a feature token for the heat conduction network, with a dimension of 1×d; , , These are the adaptive weights learned by the model, and their sum is 1. C represents the uniformity of thermally conductive phase dispersion; C represents the connectivity probability of the thermally conductive network; and O represents the fiber orientation consistency.
[0013] Furthermore, in the model pre-training step, the expression for the core coupling constraint loss function is: ,in, Loss due to core coupling constraints; Predicting losses in thermal conductivity; Predicting losses for rheological properties; For physical constraint weights; For the connectivity of the heat conduction network; The viscosity-temperature constitutive relationship of asphalt; The dynamic viscosity at the construction temperature; This refers to the construction temperature.
[0014] Furthermore, in the model pre-training step, the weakly supervised distillation is guided by weakly labeled learning generated using a traditional empirical model, which is an effective medium approximation model of thermal conductivity and filler volume fraction.
[0015] Furthermore, in the multi-task prediction step, the thermal conductivity performance indicators include thermal conductivity, thermal diffusivity, and anisotropy ratio; the rheological performance indicators include complex shear modulus, phase angle, recovery rate, non-recoverable compliance, and PG classification; the workability indicator is the viscosity range at the target temperature; the storage stability indicator is the stratification index or softening point difference; and the aging sensitivity indicator is the performance change rate before and after aging.
[0016] Furthermore, in the reverse design step, the rules of the four-dimensional dynamic manufacturability constraint library include the raw material dosage range, particle size distribution supply range, construction equipment shear rate window, orientation feasibility conditions, cost upper limit, and carbon emission upper limit; the constraint priority can be dynamically adjusted according to the scenarios of bridge deck, tunnel, airport, and high-temperature road surface, and the weights of thermal conductivity, low-temperature performance, cost, and carbon emission can be adjusted.
[0017] Furthermore, in the reverse design and multi-objective optimization steps, the objective of multi-objective optimization is to maximize the thermal conductivity or anisotropy ratio; the constraints include that the construction viscosity and irreversible flexibility do not exceed the threshold, the low temperature index and storage stability are not lower than the threshold, and the cost and carbon emissions do not exceed the upper limit; the output K candidate formulations and process parameters are accompanied by predicted performance, confidence level, key influencing factors and risk warnings.
[0018] Furthermore, in the closed-loop iteration step, the priority formula for selecting three-dimensional test points is: ,in, Score points based on selection priority; , , α represents the weighting coefficients, with a total of 1, and the default values are α=0.4, β=0.4, and γ=0.2; P represents the normalized thermal conductivity potential; U represents the normalized prediction uncertainty; and C represents the normalized experimental cost.
[0019] Furthermore, the contribution ranking of formulation variables in the interpretable output step is achieved through SHAP analysis or attention weight calculation; structural evidence includes thermal network connectivity, orientation, and interface compatibility risk; key peak curve segments are indicated by the correlation analysis between FTIR functional groups, MSCR curve segments, and performance; anomaly detection results include risk tracing outside data distribution and supplementary test recommendations.
[0020] On the other hand, a system for developing thermally conductive bitumen materials using a large-scale model is provided, comprising:
[0021] Data processing module: used to construct a database containing formulation process data and multimodal characterization data, and encode the data into a five-dimensional unified formulation structure performance representation vector constrained by component process spectrum image, wherein the vector contains thermal conduction network feature token;
[0022] Model pre-training module: used to build a large basic model of thermally conductive asphalt containing encoder and task head, and to train it using one or more of the following mechanisms: mask reconstruction, contrastive learning, thermal conductivity and rheological performance coupled constraint pre-training, and weakly supervised distillation;
[0023] Multi-task prediction module: used to build a multi-task prediction head based on a large model, outputting key performance indicators such as thermal conductivity, rheological properties, workability, storage stability, and aging sensitivity, and using one or more of the following strategies to output prediction uncertainty: ensemble model, Bayesian approximation, and Monte Carlo strategy.
[0024] Reverse Design and Multi-Objective Optimization Module: This module combines a four-dimensional dynamic manufacturability constraint library of raw materials, equipment costs, carbon emissions, and more. It performs reverse design of formulations and processes using one or more of the following methods: Bayesian optimization based on surrogate models, Pareto front search based on evolutionary algorithms, and sequential decision-making based on constraint satisfaction. The module outputs K candidate formulations and process parameters.
[0025] Closed-loop iteration module: used to calculate the priority of experimental site selection, select samples and return data after sample preparation and testing, and adjust model parameters through incremental update mechanism to achieve closed-loop iteration;
[0026] Interpretable output module: Used to output at least one interpretable piece of information from the following: ranking of the contribution of formulation variables, structural evidence, contribution indications of key spectral peak curve segments, and anomaly detection results.
[0027] Compared with existing technologies, this method and system for developing large-scale thermally conductive bitumen materials has the following advantages:
[0028] I. This invention enhances the physical adaptability of the model and improves the accuracy of performance prediction by using a five-dimensional unified representation of multimodal data and feature encoding of thermal conductivity networks, combined with multi-mechanism pre-training and a core coupling constraint loss function, thereby significantly shortening the development cycle of thermally conductive asphalt. Relying on multi-task prediction, dynamic manufacturability constraints, and a closed-loop iterative mechanism, it achieves precise reverse design for multiple objectives, balancing thermal conductivity, workability, cost, and carbon emission requirements, and reducing R&D costs. Coupled with interpretable output functionality, it clarifies key influencing factors and risk points, addressing the industry pain points of low efficiency, one-sided optimization, and untraceable results associated with traditional experience-based trial and error.
[0029] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description
[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0031] Figure 1 A flowchart illustrating the development method of thermally conductive bitumen materials for large-scale models;
[0032] Figure 2 This is a schematic diagram of the model pre-training and multi-task prediction process.
[0033] Figure 3 This is a schematic diagram of the reverse design and closed-loop iteration process. Detailed Implementation
[0034] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.
[0035] Example 1:
[0036] This embodiment is applicable to the development of thermally conductive asphalt materials for tunnels, such as... Figure 1 As shown, the overall development process of this invention includes six core steps, the specific steps of which are as follows:
[0037] First, a database of thermally conductive asphalt materials was constructed. The formulation and process data included the base asphalt type, thermally conductive filler type, fiber type, additive type, and dosage range of various raw materials. The process parameters covered heating temperature, shear rate, shear time, curing temperature, and curing time. The multimodal characterization data included FTIR spectra, DSR rheological curves, dynamic shear correlation data, SEM microscopic images, thermal conductivity test data, low-temperature performance test data, and storage stability test data.
[0038] The above data is encoded into a five-dimensional unified vector representing the formulation structure and performance, constrained by component process spectrum images. Component tokens correspond to raw material types and dosages; process tokens correspond to various process conditions; spectrum curve tokens are FTIR peak intensities and DSR curve feature values compressed and encoded; image tokens are vectors obtained from SEM microscopic images through feature extraction. Constraint tokens include upper limits for construction viscosity, low-temperature performance requirements, cost limits, and carbon emission limits. The vector also includes thermal conductivity network feature tokens obtained through a thermal conductivity network feature token encoding formula.
[0039] like Figure 2 As shown, a large-scale model of thermally conductive asphalt is then constructed. The encoder adopts a multi-layer Transformer architecture with hidden layers set to corresponding dimensions and the task header consisting of fully connected layers forming a multi-output structure. Four mechanisms are used for joint training: mask reconstruction, contrastive learning, thermal conductivity and rheological properties, coupled constraints, pre-training, weak supervision, and distillation. Mask reconstruction training randomly occludes some tokens and trains the model. Contrastive learning training treats samples from the same formula under different test conditions, aging states, and preparation batches as positive samples.
[0040] The pre-training of thermal conductivity and rheological properties coupled with constraints is achieved through the core coupled constraint loss function. The weakly supervised distillation uses an effective medium approximation model based on thermal conductivity and filler volume fraction to generate weak labels, which guide the model optimization learning.
[0041] A multi-task prediction head is built based on a pre-trained large model to output key performance indicators related to thermal conductivity, rheological properties, constructability, storage stability, and aging sensitivity. An ensemble model Bayesian approximate Monte Carlo strategy is used to jointly output prediction uncertainties, and the accuracy of uncertainty assessment is improved through multi-strategy fusion.
[0042] like Figure 3 As shown, the next step is the reverse design and closed-loop iteration phase. Combining a four-dimensional dynamic manufacturability constraint library of raw materials, equipment costs, and carbon emissions, the constraint library rules include raw material dosage range, particle size distribution, supply range, construction equipment shear rate window orientation, achievable conditions, cost upper limit, and carbon emission upper limit. For tunnel scenarios, the constraint priority is dynamically adjusted, with a focus on ensuring the weight of constructability and thermal conductivity.
[0043] The Pareto front search algorithm based on the surrogate model is used to jointly perform reverse engineering of the formulation and process. The optimization objective is to maximize the thermal conductivity. Corresponding constraints are set, and finally multiple candidate formulations and process parameters are output. Each candidate scheme is accompanied by key influencing factors of predictive performance confidence and risk warnings.
[0044] The selection priority score of each candidate sample is calculated based on the three-dimensional test site selection priority formula, and the samples with the highest scores are selected for sample preparation and testing. Asphalt samples are prepared according to the process parameters of the candidate schemes, and after the actual measurement of various performance indicators is completed, the measured data is fed back to the original database. The model parameters are adjusted through an incremental update mechanism, eliminating the need for full retraining and realizing a closed-loop iterative process of experimental relearning of the data model.
[0045] The output includes a ranking of the contribution of formulation variables, structural evidence, and key spectral peaks, along with anomaly detection results. The ranking of formulation variable contributions is obtained through a combination of SHAP analysis and attention weight calculation. Structural evidence includes risks related to thermal conductivity network connectivity, orientation, and interfacial compatibility. Key spectral peak contribution indicators are derived from the correlation analysis between FTIR functional group MSCR curve segments and material properties. Anomaly detection results include risk tracing outside the data distribution and suggestions for supplementary experiments.
[0046] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. A method for developing thermally conductive asphalt materials using large-scale modeling, characterized in that, The specific steps of this method are as follows: Data processing: Construct a database containing formulation process data and multimodal characterization data, and encode the data into a five-dimensional unified formulation structure performance representation vector constrained by component process spectrum image. The vector contains a heat conduction network feature token obtained through the heat conduction network feature token encoding formula. Model pre-training: Construct a large-scale basic model of thermally conductive asphalt containing an encoder and a task head, and train it using one or more mechanisms among mask reconstruction, contrastive learning, thermal conductivity and rheological property coupling constraint pre-training, and weakly supervised distillation. The coupling constraint pre-training is achieved through the core coupling constraint loss function. Multi-task prediction: A multi-task prediction head is established based on a large model to output key performance indicators such as thermal conductivity, rheological properties, workability, storage stability, and aging sensitivity. One or more of the following strategies are used: ensemble model, Bayesian approximation, and Monte Carlo strategy, to output prediction uncertainty. Reverse design and multi-objective optimization: Combining a four-dimensional dynamic manufacturability constraint library of raw materials, equipment costs, carbon emissions, etc., reverse design of formulation and process is performed through one or more of the following methods: Bayesian optimization based on surrogate model, Pareto front search based on evolutionary algorithm, and sequential decision based on constraint satisfaction. The output is K candidate formulations and process parameters. Closed-loop iteration: The selection priority is calculated based on the three-dimensional test point selection priority formula. After the sample preparation and testing are completed, the data is fed back and the model parameters are adjusted through an incremental update mechanism to achieve closed-loop iteration. Interpretable output: Output at least one interpretable piece of information from the following: ranking of the contribution of formulation variables, structural evidence, contribution indications of key spectral peak curve segments, and anomaly detection results.
2. The method for developing large-scale model-enabled thermally conductive asphalt materials according to claim 1, characterized in that, In the data processing, the five-dimensional unified formulation structure performance representation vector includes: component token, process token, spectral curve token, image token, and constraint token. The calculation formula for the thermally conductive network feature token is as follows: ,in, This is a feature token for the heat conduction network, with a dimension of 1×d; , , Adaptive weights for model learning; C represents the uniformity of thermally conductive phase dispersion; C represents the connectivity probability of the thermally conductive network; and O represents the fiber orientation consistency.
3. The method for developing large-scale model-enabled thermally conductive asphalt materials according to claim 1, characterized in that, In the model pre-training step, the expression for the core coupling constraint loss function is: ,in, Loss due to core coupling constraints; Predicting losses in thermal conductivity; Predicting losses for rheological properties; For physical constraint weights; For the connectivity of the heat conduction network; The viscosity-temperature constitutive relationship of asphalt; The dynamic viscosity at the construction temperature; This refers to the construction temperature.
4. The method for developing large-scale model-enabled thermally conductive asphalt materials according to claim 1, characterized in that, In the model pre-training step, weakly supervised distillation is a learning process guided by weak labels generated using a traditional empirical model, which is an effective medium approximation model of thermal conductivity and filler volume fraction.
5. The method for developing large-scale thermally conductive asphalt materials according to claim 1, characterized in that, In the multi-task prediction step, the thermal conductivity performance indicators include thermal conductivity, thermal diffusivity, and anisotropy ratio; the rheological performance indicators include complex shear modulus, phase angle, recovery rate, non-recoverable compliance, and PG classification; and the workability indicator is the viscosity range at the target temperature. Storage stability metrics are the stratification index or softening point difference; The aging sensitivity index is the rate of performance change before and after aging.
6. The method for developing large-scale model-enabled thermally conductive asphalt materials according to claim 1, characterized in that, In the reverse design step, the rules of the four-dimensional dynamic manufacturability constraint library include the raw material dosage range, particle size distribution supply range, construction equipment shear rate window, orientation feasibility conditions, cost upper limit, and carbon emission upper limit; the constraint priority can be dynamically adjusted according to the scenarios of bridge deck, tunnel, airport, and high-temperature road surface, and the weights of thermal conductivity, low-temperature performance, cost, and carbon emission can be adjusted.
7. The method for developing large-scale model-enabled thermally conductive asphalt materials according to claim 1, characterized in that, In the reverse design and multi-objective optimization steps, the objective of multi-objective optimization is to maximize the thermal conductivity or anisotropy ratio; the constraints include that the construction viscosity and irreversible flexibility do not exceed the threshold, the low temperature index and storage stability are not lower than the threshold, and the cost and carbon emissions do not exceed the upper limit; the output K candidate formulations and process parameters are accompanied by predicted performance, confidence level, key influencing factors and risk warnings.
8. The method for developing large-scale model-enabled thermally conductive asphalt materials according to claim 1, characterized in that, In the closed-loop iteration step, the priority formula for selecting three-dimensional test points is: ,in, Score points based on selection priority; , , is the weighting coefficient; P is the normalized thermal conductivity potential; U is the normalized prediction uncertainty; C is the normalized test cost.
9. The method for developing large-scale model-enabled thermally conductive asphalt materials according to claim 1, characterized in that, The contribution ranking of formulation variables in the interpretable output step is achieved through SHAP analysis or attention weight calculation; structural evidence includes thermal network connectivity, orientation, and interface compatibility risk; contribution of key spectral peak curve segments is indicated by correlation analysis between FTIR functional groups, MSCR curve segments, and performance; anomaly detection results include risk tracing outside data distribution and supplementary test recommendations.
10. A system for developing large-scale thermally conductive bitumen materials, the system being applicable to the method for developing large-scale thermally conductive bitumen materials according to any one of claims 1-9, characterized in that, The system includes: Data processing module: used to construct a database containing formulation process data and multimodal characterization data, and encode the data into a five-dimensional unified formulation structure performance representation vector constrained by component process spectrum image, wherein the vector contains thermal conduction network feature token; Model pre-training module: used to build a large basic model of thermally conductive asphalt containing encoder and task head, and to train it using one or more of the following mechanisms: mask reconstruction, contrastive learning, thermal conductivity and rheological performance coupled constraint pre-training, and weakly supervised distillation; Multi-task prediction module: used to build a multi-task prediction head based on a large model, outputting key performance indicators such as thermal conductivity, rheological properties, workability, storage stability, and aging sensitivity, and using one or more of the following strategies to output prediction uncertainty: ensemble model, Bayesian approximation, and Monte Carlo strategy. Reverse Design and Multi-Objective Optimization Module: This module combines a four-dimensional dynamic manufacturability constraint library of raw materials, equipment costs, carbon emissions, and more. It performs reverse design of formulations and processes using one or more of the following methods: Bayesian optimization based on surrogate models, Pareto front search based on evolutionary algorithms, and sequential decision-making based on constraint satisfaction. The module outputs K candidate formulations and process parameters. Closed-loop iteration module: used to calculate the priority of experimental site selection, select samples and return data after sample preparation and testing, and adjust model parameters through incremental update mechanism to achieve closed-loop iteration; Interpretable output module: Used to output at least one interpretable piece of information from the following: ranking of the contribution of formulation variables, structural evidence, contribution indications of key spectral peak curve segments, and anomaly detection results.