High-entropy alloy target performance reverse rapid proportioning method based on active learning
By using a closed-loop framework based on differentiable deep learning and active learning, combined with gradient backpropagation and attention mechanisms, the problem of the large design space of high-entropy alloy composition and the inability of traditional methods to reverse design is solved. This enables fast and accurate high-entropy alloy composition ratios that meet specific performance indicators and are engineering feasible.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIV OF MINING & TECH
- Filing Date
- 2026-05-09
- Publication Date
- 2026-06-09
AI Technical Summary
The multi-principal element characteristics of high-entropy alloys result in a large composition design space. Traditional methods struggle to quickly pinpoint the optimal ratio that meets specific performance indicators, and traditional machine learning models cannot perform reverse design, leading to long development cycles and high costs.
We employ a reverse design method based on differentiable deep learning, combined with an attention mechanism and an active learning closed-loop framework. Through gradient backpropagation and active learning, we construct a prediction-design-verification closed-loop optimization method to achieve fast and accurate design from target performance to component ratio.
It achieves efficient and precise high-entropy alloy composition design, shortens the R&D cycle, reduces costs, and ensures the engineering feasibility and stability of the design results.
Smart Images
Figure CN122177318A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of materials genome engineering and artificial intelligence, specifically involving a method for reverse design and rapid proportioning of high-entropy alloy components based on deep learning and active learning closed loop. Background Technology
[0002] High-entropy alloys (HEAs) break away from the conventional design concept of alloys using a single main element. By mixing five or more main elements in near-equal atomic ratios, they form a solid solution structure with a high-entropy effect, exhibiting excellent properties such as high strength and toughness, corrosion resistance, and resistance to high-temperature softening. With the continuous improvement of material performance requirements in extreme service environments such as aerospace and nuclear power engineering, high-entropy alloys have become a key research and development direction for next-generation structural materials.
[0003] However, while the multi-principal element characteristics of high-entropy alloys greatly expand the composition design space, they also bring about the typical problem of "combinatorial explosion." Taking a pentagonal system as an example, if the composition is explored in steps of 1 at.%, the number of candidate combinations reaches millions; if continuous mole fraction changes and higher-dimensional element systems are considered, the design parameters increase exponentially. The traditional design model, which relies on expert experience and "trial and error," has shown significant limitations in the face of such a huge parameter space, with inherent defects such as long development cycles, high costs, and difficulty in covering the entire composition space.
[0004] Existing computer-aided design methods mainly include CALPHAD thermodynamic calculations and traditional machine learning (such as random forests and GBDT). Although the former can provide high theoretical accuracy, the computation is extremely time-consuming, making it difficult to support large-scale screening; the latter, while achieving millisecond-level performance prediction, is essentially a forward mapping of "input composition - output performance," and cannot solve the inverse design problem of "given target performance - reverse calculation of composition ratio." Especially in high-dimensional continuous space, how to quickly lock the optimal ratio that meets specific performance indicators (such as specific hardness and formation enthalpy constraints) remains a key bottleneck restricting the engineering application of high-entropy alloys.
[0005] Against this backdrop, inverse design driven by differentiable deep learning offers a novel approach to solving this challenge. Among them, a closed-loop framework combining attention mechanisms and active learning stands out due to its unique advantages. This framework not only deeply captures the complex nonlinear interactions between multiple principal components through attention mechanisms, but also directly optimizes in the composition space using gradient backpropagation. Based on this idea, this invention constructs a closed-loop optimization method integrating "prediction-design-verification," which learns from massive composition-performance data, uncovers design patterns, and rapidly generates alloy proportions that meet target performance under physical constraints. Summary of the Invention
[0006] The purpose of this invention is to provide a rapid reverse-engineering method for high-entropy alloy composition based on active learning. This method aims to address the technical challenges in existing high-entropy alloy development, such as low screening efficiency due to "combinatorial explosion," the inability of traditional machine learning models to deduce composition ratios based on target performance, and poor model generalization ability under small sample conditions. This invention achieves rapid, accurate, and engineering-feasible composition design of high-entropy alloys driven by target performance by constructing a differentiable performance prediction model and a gradient backpropagation mechanism, combined with an active learning closed loop. This significantly shortens the material development cycle and reduces experimental costs.
[0007] To achieve the above objectives, the present invention adopts the following technical solution: a method for reverse rapid proportioning of high-entropy alloy target properties based on active learning, comprising the following steps: Step 1, Determine design requirements and establish engineering constraints: Set the composition normalization constraints, preparation process constraints and phase stability constraints that the composition vector of the high-entropy alloy must satisfy; Step 2, constructing a digital characterization and design space for high-entropy alloy composition data: set a set containing n candidate elements, and define the composition ratio of any high-entropy alloy as an n-dimensional continuous numerical vector, where each component in the vector represents the mole fraction of the corresponding element, thereby constructing a continuous design space; Step 3, Construct and pre-train a differentiable performance predictor based on attention mechanism: Construct a deep neural network with attention mechanism as a performance predictor, and pre-train it using the initial sample set to establish a differentiable mapping from component vectors to target performance; Step 4, perform component inverse design based on gradient backpropagation: fix the pre-trained performance predictor, construct a composite loss function driven by the target performance, and iteratively update the component vector initialized in the design space through gradient optimization algorithm. At the same time, force the component normalization constraint to be satisfied during the iteration until the theoretically optimal candidate component is output. Step 5, introduce active learning closed loop for iterative optimization: screen and verify the candidate components output in Step 4, expand the real performance data obtained from the verification to the training set to update the performance predictor, and repeat Step 4 to form a closed loop optimization process.
[0008] Furthermore, in step 1, the engineering constraints include: Component normalization constraint: Component vector of any design scheme The mole fraction normalization condition must be met, i.e. And the mole fraction of a single element ; Preparation process constraints: Set a minimum component adjustment step size and a minimum addition threshold for a single element; Phase stability constraints: The design process must satisfy the valence electron concentration (VEC) or enthalpy of mixing. Specific thermodynamic threshold constraints.
[0009] Furthermore, in step 2, a candidate element set is set. The component ratio is defined as a 1×n-dimensional continuous numerical vector. ; where the i-th channel of the vector corresponds to the i-th element in the candidate element set, and the value of that channel is... Mole fractions representing the corresponding elements; each mole fraction The value range is preset. , Represents mole fraction The minimum value, Represents mole fraction The maximum value that can be obtained.
[0010] Furthermore, in step 3, the deep neural network sequentially includes: The element embedding layer is used to map each element in the input component vector into a feature vector containing physical and chemical properties such as atomic radius, electronegativity, and valence electron number; A multi-head self-attention encoding layer is used to capture nonlinear interactions between elements and generate a global feature vector of the alloy. The performance output layer maps the global feature vector of the alloy to the predicted value of the target performance through a multi-layer fully connected network.
[0011] Furthermore, in step 3, the construction and training of the performance predictor are based on the following setting: the performance of high-entropy alloys is determined by their chemical composition and the interaction between elements, ignoring the randomness of the microstructure in the initial composition design stage.
[0012] Furthermore, in step 4, the composite loss function Loss is defined as:
[0013] Among them: the first item is the performance target-driven item, which guides components toward the target performance. Approaching; This is a performance prediction value. The target performance value; Second item This is a physical stability constraint term used to penalize components that do not meet single-phase stability requirements; Third item These are engineering constraints used to determine the mole fraction. Limited to Within the range; α, β, and γ are weighting coefficients.
[0014] Furthermore, the physical stability constraint term Based on valence electron concentration (VEC), its specific form is as follows:
[0015] in, The preset VEC threshold, It is a linear rectifier function.
[0016] Furthermore, in step 4, the component vectors initialized within the design space are randomly generated in multiple sets using a Gaussian or uniform distribution.
[0017] Furthermore, in step 4, the gradient descent algorithm is used to process the component vector. Perform iterative updates, and immediately after each update, perform a normalization projection operation to force the updated vector to satisfy the constraint that the sum of the mole fractions of all elements is 1.
[0018] Furthermore, in step 5, the basis for screening and verifying samples is the uncertainty estimate or expected improvement of the performance prediction performance of the performance predictor for candidate components; the high-fidelity verification method is experimental preparation and performance testing, or physical simulation based on first-principles calculation and phase diagram calculation.
[0019] Beneficial effects: Compared with the prior art, the present invention has the following advantages: 1. This invention achieves true reverse engineering for target performance, overcoming the limitations of traditional trial-and-error methods and forward prediction. Utilizing the differentiability of deep neural networks, it directly calculates the gradient of performance indicators relative to the composition vector through gradient backpropagation. This eliminates the need for brute-force traversal or random searches of massive amounts of composition in alloy design. Instead, it directly derives the theoretically optimal composition ratio based on given mechanical / physical performance targets, fundamentally solving the problem that traditional machine learning methods cannot perform reverse engineering.
[0020] 2. Significantly improves optimization efficiency and global search capability in high-dimensional continuous component spaces, solving the problems of "combinatorial explosion" and local optima. Addressing the shortcomings of traditional discrete optimization algorithms such as genetic algorithms and particle swarm optimization, which are prone to getting trapped in local optima and have slow convergence in high-dimensional spaces, this invention combines a gradient continuous optimization mechanism with normalized projection operations. While ensuring the physical meaning of the components, it significantly reduces computational load and search time, achieving millisecond-level component generation. To solve the dependency of gradient algorithms on initial values, a multi-starting-point random initialization strategy based on Gaussian or uniform distribution is further introduced. Within the constraints, multiple sets of initial component vectors distributed in different regions are generated as the starting positions for gradient descent, effectively avoiding the local optimum trap caused by a single initial value and significantly improving the probability and stability of the algorithm locking the global optimum in complex component spaces.
[0021] 3. An active learning closed loop of "prediction-design-verification" is constructed, effectively overcoming the challenges of data scarcity and model generalization. Addressing the high cost of acquiring experimental data for high-entropy alloys, this invention introduces an active learning mechanism. High-value samples (those with high uncertainty or high expected improvement) are selected for verification, and the model is incrementally updated. This approach allows the model to "learn while designing," continuously improving its accuracy in key design regions with each iteration, thus ensuring reliable design results even with small sample sizes.
[0022] 4. By integrating physical metallurgical rules and engineering preparation constraints, the engineering feasibility of the design scheme is guaranteed. This invention does not rely solely on data-driven approaches, but explicitly embeds phase stability criteria such as valence electron concentration (VEC) and enthalpy of mixing, as well as the batching accuracy requirements of the smelting process, into the loss function and optimization process. This ensures that the generated alloy proportions not only meet theoretical performance standards but also form stable single-phase solid solutions, satisfying the process requirements of actual metallurgical preparation and avoiding the generation of ineffective or unpreparable composition schemes. Attached Figure Description
[0023] Figure 1 This is a schematic diagram illustrating the entire process and optimization principle of the high-entropy alloy composition reverse rapid proportioning method based on gradient backpropagation in this invention. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be clearly and completely described below in conjunction with the specific description in the technical disclosure. Obviously, the described embodiments are only a part of the embodiments of this invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.
[0025] This invention provides a rapid reverse-engineering method for high-entropy alloy target performance based on active learning. Its core lies in constructing an active learning closed loop of "prediction-design-verification," achieving rapid and accurate reverse design from target performance to component proportioning through a differentiable performance prediction model and gradient backpropagation mechanism. The specific steps are described in detail below. The technical solution of this invention is mainly based on the following three interrelated core mechanisms: First, a general framework for inverse design based on the fusion of gradient backpropagation and physical constraints was constructed. Through a "differentiable predictor—inverse designer—active learning closed-loop" architecture, the high-dimensional combinatorial design problem is transformed into a gradient optimization problem in continuous space. This framework overcomes the limitation of traditional forward prediction models being unable to perform inverse design by freezing the pre-trained predictor, directly calculating the gradient of the target performance with respect to the component vector using automatic differentiation, and combining normalized projection to inversely deduce the optimal components.
[0026] Second, a component feature extraction mechanism integrating attention mechanism and a small-sample closed-loop optimization mechanism based on active learning is proposed. A multi-head self-attention mechanism is introduced into the predictor to accurately capture the complex nonlinear interactions ("cocktail effect") between multiple principal components. At the same time, an active learning closed loop of "design-validation-incremental update" is constructed. By selecting high-value samples for validation and feedback, the model is driven to continuously improve itself under small-sample conditions, significantly improving design efficiency and accuracy.
[0027] Third, a composite loss function guidance mechanism incorporating physical metallurgical rules and engineering preparation constraints was established. Phase stability criteria such as valence electron concentration (VEC) and enthalpy of mixing, as well as preparation accuracy requirements, were explicitly constructed as physical and engineering constraint terms in the loss function. This mechanism guides the optimization process to generate component ratios that simultaneously satisfy target performance, thermodynamic stability, and process feasibility, ensuring the engineering usability of the design results.
[0028] The following combination Figure 1 The schematic diagram showing the entire process of the method of the present invention provides a detailed description of the invention. The present invention provides a method for reverse rapid proportioning of high-entropy alloy target properties based on active learning, comprising the following steps:
[0029] Step 1: Determine design requirements and establish engineering constraints The method proposed in this invention follows the following engineering and physical constraints: (1) Component normalization constraint: The component vector of any design scheme The mole fraction normalization condition must be met, i.e. And the mole fraction of a single element .
[0030] (2) Preparation process constraints: To adapt to the actual batching precision of vacuum arc melting or powder metallurgy, a minimum component adjustment step size (e.g., 0.1 at.%) is set, and a minimum addition threshold is set for each single element (e.g., ...). This is to avoid problems such as difficulty in controlling trace elements or excessive segregation of principal components.
[0031] (3) Phase stability constraints: To ensure that the designed alloy is a single-phase solid solution (FCC or BCC) rather than a brittle intermetallic compound, the design process must meet the requirements of valence electron concentration (VEC) or enthalpy of mixing ( ). ) specific thermodynamic threshold constraints (e.g., for the FCC phase, constraints) );
[0032] Data characterization assumptions: It is assumed that the performance of high-entropy alloys is mainly determined by their chemical composition and the interaction between elements, and the randomness of microstructure (grain size, defects) in the initial composition design stage is ignored.
[0033] Step 2: Constructing a digital characterization and design space for high-entropy alloy composition data To achieve continuous gradient optimization of composition, a digital characterization matrix for the high-entropy alloy composition must first be established. A candidate element set is then defined. (For example The composition ratio of any high-entropy alloy is defined as follows: Continuous numerical vectors The data input format is shown in Table 1 below. This structure will serve as the direct input interface for subsequent deep neural networks.
[0034] Table 1 Data Input Format
[0035] The above definition transforms the discrete combinatorial problem into a vector optimization problem in a continuous space, laying the foundation for subsequent gradient calculation.
[0036] Step 3: Construct and pre-train an attention-based differentiable performance predictor This step aims to establish a continuously differentiable mapping model from "component vectors" to "target performance".
[0037] (1) Model construction: Construct a deep neural network with self-attention mechanism as a performance predictor.
[0038] Element embedding layer: maps the input element channel indices to high-dimensional feature vectors containing physicochemical properties (such as atomic radius, electronegativity, and number of valence electrons).
[0039] Attention Encoding Layer: Utilizes a multi-head self-attention mechanism to capture the complex nonlinear interactions (i.e., the "cocktail effect") between different element pairs (such as Co-Cr, Ni-Al) in high-entropy alloys, generating a global feature vector for the alloy. Performance Output Layer: A multi-layer fully connected network (MLP output layer) maps the global feature vector of the alloy to the target performance prediction value. (e.g., Vickers hardness HV).
[0040] (2) Model pre-training: Construct an initial sample set using existing high-throughput computing data (such as CALPHAD, DFT) or public databases. By minimizing prediction performance Compared to real performance The mean squared error (MSE) between the network weights is used to optimize the network weight parameters. The loss function is defined as follows: ; After training, all weight parameters of the predictor are fixed to establish the initial performance response surface, providing a fixed gradient calculation basis for subsequent reverse design.
[0041] Step 4: Construct a component inverse designer based on gradient backpropagation This step is the core of the invention. As a "reverse model," it calculates the gradient through backpropagation based on the mapping relationship learned by the predictor, and optimizes from a stochastic initial design to obtain a high-performance material. The specific implementation includes the following steps: (1) Constructing a differentiable environment for the transfer predictor parameters: All weight parameters of the performance predictor trained in step 3 are transferred to the differentiable environment. Fix and assign it to the designer, treating it as a fixed, differentiable performance function. At this point, the designer no longer updates the network parameters; it only updates the network's input (component vector). ) is the only new learning variable.
[0042] (2) Generating random initial design schemes: In order to avoid the design getting trapped in local optima due to a single initial value, multiple sets of initial component vectors are randomly generated within the constraints using a Gaussian distribution or a uniform distribution. These initial points are distributed in different regions of the component space, serving as the starting points for gradient descent and thus increasing the probability of finding the global optimum.
[0043] (3) Optimize design variables through backpropagation
[0044] Constructing a composite loss function: To ensure that the generated components satisfy both performance metrics and physical / engineering constraints, a total loss function to be minimized is defined. as follows: ; Among them: the first item is the performance target-driven item, which guides components toward the target performance. Approaching; This is a performance prediction value. The target performance value; Second item This is a physical stability constraint term used to penalize components that do not meet single-phase stability requirements; Third item These are engineering constraints used to determine the mole fraction. Limited to Within the range; α, β, and γ are weighting coefficients.
[0045] Among them, physical stability constraint term Based on valence electron concentration (VEC), its specific form is as follows: ; in, The preset VEC threshold, It is a linear rectifier function.
[0046] Gradient inverse optimization: Utilizing the automatic differentiation mechanism (chain rule) of deep learning frameworks, the total loss function is calculated in one step with respect to the input component vector. analytic gradient That is, to calculate the loss for each element component in the component vector. Partial derivatives: ; Iterative update: Using gradient descent algorithms (such as Adam), the allocation ratio is updated based on the calculated gradient direction. ; Key operation: Immediately after each gradient update, perform... Perform boundary trimming to ensure that the content of each element is within the specified range. Within the range, and normalized projection, force it to satisfy... The physical conditions. After After the rounds of iteration converge, the theoretically optimal candidate components are output. .
[0047] Step 5: Closed-loop iteration and incremental model update based on active learning To address the problem of weak extrapolation ability of models caused by the scarcity of high-entropy alloy data, an active learning strategy is introduced to establish a closed loop of "prediction-design-verification-feedback".
[0048] (1) Screening of high-value samples: From the candidate component set generated in step 4, the samples with the highest validation value are screened using the uncertainty estimate or expected lift of the model. .
[0049] (2) High-fidelity verification and data augmentation: Experimental preparation (vacuum arc melting + hardness testing) or high-fidelity physical simulation (DFT / CALPHAD) is performed on the selected samples to obtain their true performance values. To eliminate model prediction bias, the new data is compared. Added to the original training set to form an augmented dataset. .
[0050] (3) Predictor fine-tuning: using augmented datasets The performance predictor is then retrained or fine-tuned. The updated predictor will have higher local prediction accuracy, and then proceed to the next round of reverse design in step 4. Through the above closed-loop iteration, as validation data accumulates, the design results will gradually approach the true physical optimum, ultimately achieving rapid and accurate determination of the high-entropy alloy composition.
[0051] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A rapid reverse proportioning method for high-entropy alloy target properties based on active learning, characterized in that: Includes the following steps: Step 1, Determine design requirements and establish engineering constraints: Set the composition normalization constraints, preparation process constraints and phase stability constraints that the composition vector of the high-entropy alloy must satisfy; Step 2, constructing a digital characterization and design space for high-entropy alloy composition data: set a set containing n candidate elements, and define the composition ratio of any high-entropy alloy as an n-dimensional continuous numerical vector, where each component in the vector represents the mole fraction of the corresponding element, thereby constructing a continuous design space; Step 3, Construct and pre-train a differentiable performance predictor based on attention mechanism: Construct a deep neural network with attention mechanism as a performance predictor, and pre-train it using the initial sample set to establish a differentiable mapping from component vectors to target performance; Step 4, perform component inverse design based on gradient backpropagation: fix the pre-trained performance predictor, construct a composite loss function driven by the target performance, and iteratively update the component vector initialized in the design space through gradient optimization algorithm. At the same time, force the component normalization constraint to be satisfied during the iteration until the theoretically optimal candidate component is output. Step 5, introduce active learning closed loop for iterative optimization: screen and verify the candidate components output in Step 4, expand the real performance data obtained from the verification to the training set to update the performance predictor, and repeat Step 4 to form a closed loop optimization process.
2. The method according to claim 1, characterized in that: In step 1, the engineering constraints include: Component normalization constraint: Component vector of any design scheme The mole fraction normalization condition must be met, i.e. And the mole fraction of a single element ,in, Indicates the channel index. Indicates the first Mole fraction of each channel; Preparation process constraints: Set a minimum component adjustment step size and a minimum addition threshold for a single element; Phase stability constraints: The design process must satisfy the valence electron concentration (VEC) or enthalpy of mixing. Specific thermodynamic threshold constraints.
3. The method according to claim 1, characterized in that: In step 2, a candidate element set is set. ,in, Represent candidate elements; define the proportion as a 1×n dimensional continuous numerical vector. ; where the vector's th The channel corresponds to the first candidate element in the set of candidate elements. A type of element, the value of that channel. Mole fractions representing the corresponding elements; each mole fraction The value range is preset. , Represents mole fraction The minimum value, Represents mole fraction The maximum value that can be obtained.
4. The method according to claim 1, characterized in that: In step 3, the deep neural network includes, in sequence: The element embedding layer is used to map each element in the input component vector into a feature vector containing physical and chemical properties such as atomic radius, electronegativity, and valence electron number; A multi-head self-attention encoding layer is used to capture nonlinear interactions between elements and generate a global feature vector of the alloy. The performance output layer maps the global feature vector of the alloy to the predicted value of the target performance through a multi-layer fully connected network.
5. The method according to claim 1, characterized in that: In step 3, the construction and training of the performance predictor are based on the following setting: the performance of high-entropy alloys is determined by their chemical composition and the interaction between elements, ignoring the randomness of the microstructure in the initial composition design stage.
6. The method according to claim 3, characterized in that: In step 4, the composite loss function Loss is defined as: Among them: the first item is the performance target-driven item, which guides components toward the target performance. Approaching; This is a performance prediction value. The target performance value; Second item This is a physical stability constraint term used to penalize components that do not meet single-phase stability requirements; Third item These are engineering constraints used to determine the mole fraction. Limited to Within the range; α, β, and γ are weighting coefficients.
7. The method according to claim 6, characterized in that: The physical stability constraint Based on valence electron concentration, its specific form is as follows: ； in, The preset VEC threshold, It is a linear rectification function.
8. The method according to claim 1, characterized in that: In step 4, the component vectors initialized in the design space are randomly generated in multiple sets using a Gaussian or uniform distribution.
9. The method according to claim 1 or 6, characterized in that: In step 4, the gradient descent algorithm is used to iteratively update the component vector x. After each update, a normalization projection operation is immediately performed to force the updated vector to satisfy the constraint that the sum of the mole fractions of all elements is 1.
10. The method according to claim 1, characterized in that: In step 5, the basis for screening and verifying samples is the uncertainty estimate or expected improvement of the performance prediction of candidate components by the performance predictor; the high-fidelity verification method is experimental preparation and performance testing, or physical simulation based on first-principles calculation and phase diagram calculation.