LLM-GA collaborative optimization and hybrid coding-based high-entropy alloy multi-element combination screening method and system

Abstract

The invention discloses a high-entropy alloy multi-element combination screening method and system based on large language model (LLM)-genetic algorithm (GA) collaborative optimization and hybrid coding, and the system comprises a literature intelligent analysis module, a genetic algorithm optimization module and an experimental verification module, extracting element activity characteristics from the multi-source literature; a genetic algorithm (GA) optimization module adopts a binary-real number hybrid coding mechanism, and candidate metal combinations are generated through collaborative optimization of element type selection and proportional distribution; and the experiment verification module verifies optimization results in batches through a high-throughput experiment interface, and feeds back experiment data to the knowledge base and the algorithm parameter library. Through deep cooperation of the LLM and the genetic algorithm, the research and development period of a traditional high-entropy alloy is shortened to the hour level from the thousand level, and the design efficiency of a multi-element catalyst is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the interdisciplinary technical field of material design and artificial intelligence, and specifically relates to a method and system for intelligent screening of high-entropy alloy multi-element combinations based on LLM-GA collaborative optimization and hybrid coding. Background Art

[0002] High-entropy alloys (HEAs), as groundbreaking multifunctional materials, have demonstrated revolutionary potential in catalytic reactions (such as hydrogen evolution reactions) and energy storage (such as fuel cell electrodes) due to their unique chemical disorder and multi-element synergistic effects. However, their development faces two major fundamental scientific challenges: First, the number of candidate combinations for five or more elements in an alloy increases exponentially due to the combinatorial explosion effect. When the number of candidate elements N ≥ 20, the theoretical number of combinations exceeds 10. 12 The traditional trial-and-error method would take thousands of years even at a rate of 100 experiments per day. Secondly, cross-scale optimization requires the integration of literature knowledge (such as the intrinsic activity of platinum group metals) and experimental data (such as the effect of lattice distortion on stability), but existing technologies cannot effectively integrate these two types of heterogeneous data sources.

[0003] Current mainstream technologies have significant limitations. In terms of single-method dimensions: ① First-principles calculations (such as VASP software) are limited by computational complexity and can only process alloy systems with ≤3 elements; ② Machine learning methods (such as random forest regression) rely on 10 4 3. While genetic algorithms (GAs) can handle multi-element optimization, their randomly generated initial populations result in slow convergence. Knowledge gaps further exacerbate the challenge: ① Existing element screening methods rely solely on keyword frequency statistics in literature, ignoring the rules governing synergistic effects between elements; ② Over 92% of published literature data is not structured, leaving potential association rules unexplored; and ③ New research results require an average of 6-12 months of manual integration, resulting in a lag in knowledge updates.

[0004] Experimental validation also faces multiple bottlenecks: ① The traditional single-sample synthesis-characterization process takes over 48 hours, failing to meet high-throughput requirements; ② Existing automated experimental platforms (such as MIT's autonomous laboratory) lack real-time data exchange protocols with computational modules, resulting in feedback delays exceeding 24 hours; ③ In multi-objective optimization scenarios, experimental data utilization is less than 35%, with a significant amount of data not being used to update predictive models. These technical limitations collectively hinder the efficient development of high-entropy alloys and highlight the urgent need to establish a multi-source data fusion and closed-loop optimization system. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention provides a method and system for intelligent screening of high-entropy alloy multi-element combinations based on LLM-GA collaborative optimization and hybrid coding. This method uses a large language model to construct a dynamic knowledge graph to guide genetic algorithm optimization. Binary-real hybrid coding is used to achieve collaborative optimization of element selection and ratios. High-throughput experimental feedback is combined to form a closed-loop iterative mechanism. This shortens the traditional R&D cycle from millennia to hours.

[0006] Specifically, the method includes the following steps:

[0007] Step 1: Count the number of occurrences of various metal elements in all high-entropy alloy-related literature, sort them from most to least, and select the top n metal elements as the most frequently occurring metal elements. Relevant literature refers to the literature found by entering the keyword "high-entropy alloy" in the literature database. This means that the literature contains "high-entropy alloy" in the literature content.

[0008] Step 2: Based on the metal elements screened in step 1, a hybrid coding genetic algorithm is used to generate candidate metal combinations; the hybrid coding genetic algorithm includes a binary segment and a real segment; the binary segment is an element selection segment, and the real segment is a proportion allocation segment;

[0009] Step 3: Based on the candidate metal combinations generated in step 2, synthesize candidate alloys and test their properties;

[0010] Step 4: Use the candidate alloys as individuals to construct a population. Based on the performance test results of step 3, use the fitness function to rank the alloys in the current population by performance, and eliminate individuals ranked in the bottom [20%] of the performance ranking, where [20%] refers to the integer value after rounding off. The fitness function is designed as F = 0.6 catalytic activity + 0.3 stability - 0.1 * precious metal cost;

[0011] Step 5: Generate new metal combinations through hybrid coding genetic algorithm;

[0012] Step 6: Iterate steps 2 to 5 until the individual with the best performance is obtained, that is, the alloy with the best metal combination performance is obtained.

[0013] Furthermore, the chromosome structure of the hybrid coding genetic algorithm includes an element selection segment and a proportion distribution segment, wherein the binary segment Ei∈{0,1} is used to identify the selection status of the element, 1 is selected and 0 is not selected; the real number segment Pi∈[0.05,0.35] represents the atomic percentage of each element, and satisfies the condition ΣPi=1, and the sum of the proportions of all elements is 100%.

[0014] The present invention also provides a high entropy alloy multi-element combination screening system based on LLM-GA collaborative optimization and hybrid coding, comprising:

[0015] Intelligent Document Analysis Module: This module builds a dynamic knowledge base from multi-source documents using a large language model (LLM), constructs a document semantic network based on the Transformer architecture, and generates element activity feature vectors using a multi-head attention mechanism.

[0016] In some embodiments of the present invention, a five-dimensional activity feature vector (Pt, Pd, Rh, Ir, Ru) can be output, automatically and incrementally training and updating the element synergy rules in the knowledge graph every week. Through intelligent analysis of extensive literature, this module can effectively identify the activity of each element and its synergistic effects under different conditions, providing strong theoretical support for alloy design.

[0017] Genetic algorithm optimization module: uses binary-real mixed coding chromosomes, including element selection segments and proportion distribution segments, and implements a segmented crossover mutation strategy;

[0018] Experimental verification module: Integrates high-throughput synthesis equipment with performance testing units and establishes a data feedback channel in JSON-LD format;

[0019] Dynamic Knowledge Management Subsystem: This subsystem builds an element-performance association graph based on a graph database, supports multimodal associations between literature feature vectors and experimental data, and enables incremental updates of the knowledge base through version control. This subsystem efficiently manages large amounts of data from diverse sources and ensures that the knowledge and data in the system remain up-to-date, adapting to evolving research needs.

[0020] Furthermore, the chromosome encoding structure of the genetic algorithm optimization module is [E1, E2, ..., En | P1, P2, ..., Pn], where:

[0021] Ei∈{0,1} represents the element selection state, Pi∈[0.05,0.35] is the atomic percentage and satisfies ΣPi=1;

[0022] The fitness function is F = 0.6 catalytic activity + 0.3 stability - 0.1 * precious metal cost.

[0023] Furthermore, the working steps of the genetic algorithm optimization module include:

[0024] (1) Crossover operation is performed in segments: uniform crossover operation is used for binary coding segments; simulated binary crossover operation is used for real coding segments;

[0025] (2) The mutation operation is implemented in a hierarchical manner: bit flip mutation is performed on the binary segment; polynomial mutation is performed on the real segment; this strategy can improve the global search capability of the genetic algorithm and ensure diversity in the search process, avoiding falling into local optimal solutions.

[0026] (3) Population elimination mechanism: With candidate alloys as individuals, a population is constructed. Based on the performance test results, the metal combinations in the current population are ranked by performance according to the preset fitness function, and the individuals ranked in the bottom 20% are eliminated. The number of eliminated combinations is an integer value rounded off to 20% of the total number of the current population. The fitness function is designed as F = 0.6 catalytic activity + 0.3 stability - 0.1 * precious metal cost; the number of eliminated combinations

[0027] (4) Large language model recommendation injection: The element ratios, performance data, and real-time parameters of the experimental verification module of the remaining metal combinations are input into the large language model to generate new element combinations that are at least the same as the number of eliminated ones;

[0028] (5) Hybrid population iteration: The new element combinations recommended by the large language model are mixed with the retained original population to form a new generation of population for iterative optimization. This optimization module can quickly find high-performance, high-stability, and cost-effective high-entropy metal combinations within a limited number of iterations, and has strong practicality.

[0029] Furthermore, the update rule of the dynamic knowledge base is: K_new = αK_old + (1-α) * Σφ(x_i) / n; where α=0.85 is the knowledge decay factor, and φ(·) is the feature mapping function based on GAT.

[0030] The dynamic knowledge base is constructed by extracting <element, performance parameter, test condition> triplets from the literature. The inter-element synergy coefficient is calculated using the GAT graph attention network. When new experimental data is imported, the following update rule is applied: K_new = αK_old + (1-α) * Σφ(x_i) / n, where α = 0.85 is the knowledge decay factor and φ(·) is the feature mapping function. This method enables real-time updating of the knowledge base, ensuring that alloy design is always based on the latest research results and experimental data, thereby improving prediction accuracy.

[0031] Furthermore, the experimental verification module supports parallel processing of multiple samples / batches and includes: an ultrafast heating device for achieving rapid temperature increase processing of materials, including one of a high-temperature thermal shock device, a laser heating device, or a microwave heating device, wherein the heating rate of the high-temperature thermal shock device is ≥100°C / s;

[0032] Structural analysis equipment for monitoring microstructural changes in materials, including one or more of an in-situ X-ray diffraction (XRD) unit, a Raman spectrometer, or an infrared spectrometer;

[0033] Performance testing equipment used to evaluate the functional properties of materials, including one or more of an electrochemical workstation, a mechanical strength tester, or a thermal conductivity tester;

[0034] The experimental verification module transmits real-time data back to the genetic algorithm optimization module via JSON-LD format, triggering dynamic adjustment of the genetic algorithm parameters. (When the error exceeds 10%, the mutation rate increases by 0.1.) This module can efficiently synthesize a large number of alloy samples and conduct real-time performance evaluation, providing real-time feedback for screening and optimizing metal combinations.

[0035] Furthermore, the data processing steps of the experimental verification module are:

[0036] (1) Synthesis parameter standardization: discretize the atomic percentage of the elements in steps of 0.05;

[0037] (2) Abnormal data filtering: automatically remove sample data with XRD crystallinity lower than 80%;

[0038] (3) Feature coding processing: Use one-hot coding to characterize the alloy composition combination; extract electronic structure features through RDKit molecular descriptors; these data processing methods can ensure the high quality of experimental data and provide reliable data support for subsequent algorithm optimization.

[0039] (4) Encapsulate the processed data in JSON-LD format and send it back.

[0040] Furthermore, the system features a visual monitoring interface: a primary curve dynamically displays the evolution of the optimal fitness of the genetic algorithm population; a secondary curve displays the Shannon diversity index in parallel; a three-dimensional scatter plot maps the relationship between the catalytic activity and stability of element combinations and the cost of precious metals; and an interactive control panel allows real-time adjustment of the elite retention ratio (5%-15%), crossover rate (0.7-0.9), and mutation rate (0.01-0.05). This interface collaboratively presents the optimization process through multi-dimensional visualization, enabling dynamic intervention in algorithm parameters. This interface allows users to clearly monitor the optimization process and adjust parameters in a timely manner to ensure efficient system operation and achieve optimal optimization results.

[0041] The benefits of the present invention lie in: through the LLM-GA collaborative architecture and hybrid coding mechanism, the R&D cycle of high-entropy alloys with five elements or more is shortened from the millennia of traditional trial-and-error methods to less than 72 hours, significantly improving the efficiency of element combination screening; the dynamic knowledge base integrates implicit rules from the literature to promote the improvement of the mean fitness of the initial population of the genetic algorithm, and combined with a high-throughput closed-loop experimental verification system, it achieves a reduction in the amount of precious metals used and an improvement in catalytic stability. The utilization rate of experimental data is greatly improved, and the comprehensive development cost is reduced to 0.7% of the traditional method. The present invention successfully overcomes the problems of combinatorial explosion and cross-scale optimization, providing a disruptive technical path for the design of clean energy catalysts. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is the main interface of the system;

[0043] Figure 2 for the literature screening process;

[0044] Figure 3 is the element frequency statistics;

[0045] Figure 4 Schematic diagram of high-throughput preparation;

[0046] Figure 5 The catalyst overpotential distribution during the high-throughput experiment;

[0047] Figure 6 The genetic algorithm was validated in 20 independent experiments. DETAILED DESCRIPTION

[0048] The following examples are used to further illustrate the present invention. Their purpose is to illustrate the present invention and should not be construed as limiting the scope of the present invention. Unless otherwise specified, all references are by weight and weight percentage.

[0049] Unless otherwise specified, the raw materials used in the present invention are conventional commercial products; the methods used in the present invention are conventional methods in the art unless otherwise specified.

[0050] The relevant literature referred to in this invention is obtained by searching a literature database by entering the keyword "high entropy alloy." This refers to literature containing the term "high entropy alloy" in its content. Literature databases commonly used in this field include, but are not limited to, China National Knowledge Infrastructure (CNKI), Wanfang Data Knowledge Service Platform, and Web of Science.

[0051] The embodiments of the present invention are further described below with reference to a number of embodiments.

[0052] It should be clear that the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0053] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. The singular forms "a", "an", "the" and "the" used in the embodiments of the present application and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise.

[0054] Example 1

[0055] During the element pre-screening phase, 90 metallic elements were screened using the InternLM 2.5 7B large language model (LoRA fine-tuned). Ten elements with high potential were ultimately selected: Pt, Ni, Ir, Pd, Cu, Mn, Fe, Co, Zn, and Ru. These elements were identified based on the criteria of a literature frequency greater than 5% and an element synergy score greater than 0.7. This process reduced the number of candidate combinations from 43,949,268 to 126.

[0056] In the hybrid encoding design, the chromosome structure consists of binary segments and real number segments. The binary segment Ei∈{0,1} is used to identify the selection state of the elements (for example, [1,1,1,1,1] indicates the selection of Pt, Ni, Ir, Pd, and Cu). The real number segment Pi∈[0.05,0.35] represents the atomic percentage of each element, and satisfies the condition ΣPi=1. The step size is 0.05 to ensure that the total percentage of all elements is 100%.

[0057] In the iterative optimization phase, six candidate alloys were synthesized in each batch, and the overpotential (10 mA / cm²) of the hydrogen evolution reaction (HER) was obtained through high-throughput electrochemical testing.

[0058] At this time, the genetic algorithm adopts a directed replacement strategy by eliminating the 20% individuals with the lowest fitness and injecting the element combination recommended by LLM (such as PtNiIrPdRh). After 4 iterations, it finally locks in the optimal combination, NiCuPdIrPt, with an overpotential of 25.5mV.

[0059] In the preferred scheme, the generation of the initial population is constrained by the element activity feature vector output by the LLM (for example, Pt=0.92, Ni=0.78, Ir=0.65), ensuring that 50% of the individuals contain the top three high-scoring elements. In addition, the fitness function includes catalytic activity prediction (weight 0.6), stability prediction (weight 0.3), and precious metal cost (weight -0.1), with the goal of optimizing catalytic performance, stability, and cost.

[0060] Table 1 Overpotentials of 126 element combinations

[0061] Sample ID Elements Voltage / mV Sample ID Elements Voltage / mV a 0 Fe Co Cu Ir Pt 49.4 c 11 Mn Fe Ni Ru Pt 37.3 a 1 Mn Fe Co Cu Pt 52 c 12 Mn Fe Ni Zn Pt 42.5 a 2 Fe Co Ni Cu Pt 59 c 13 Fe Ni Ru Pd Pt 38.3 a 3 Fe Co Cu Pd Pt 56.8 d 0 Fe Ni Zn Pd Pt 55.6 a 4 Fe Co Cu Zn Pt 50.7 d 1 Fe Ni Zn Ru Pt 36.1 a 5 Mn Co Cu Ir Pt 71.2 d 2 Mn Ni Pd Ir Pt 37.4 a 6 Co Ni Cu Ir Pt 60.3 d 3 Mn Ni Ru Ir Pt 46.9 a 7 Co Cu Pd Ir Pt 147.8 d 4 Mn Ni Zn Ir Pt 44.6 a 8 Co Cu Ru Ir Pt 47.8 d 5 Mn Ru Pd Ir Pt 39.9 a 9 Co Cu Zn Ir Pt 48.8 d 6 Ni Ru Pd Ir Pt 58.5 a 10 Mn Co Ni Cu Pt 52.6 d 7 Ni Zn Pd Ir Pt 36.9 a 11 Mn Co Cu Pd Pt 49.1 d 8 Ni Zn Ru Ir Pt 54.2 a 12 Mn Co Cu Zn Pt 53 d 9 Zn Ru Pd Ir Pt 41 a 13 Co Ni Cu Pd Pt 54.8 d 10 Mn Ni Ru Pd Pt 53.1 b 0 Co Ni Cu Ru Pt 62 d 11 Mn Ni Zn Pd Pt 51.5 b 1 Co Ni Cu Zn Pt 51.9 d 12 Mn Ni Zn Ru Pt 35.3 b 2 Co Cu Ru Pd Pt 76.6 d 13 Mn Fe Cu Ir Pt 52.3 b 3 Co Cu Zn Pd Pt 51.7 e 0 Mn Fe Cu Zn Pt 41.6 b 4 Mn Fe Co Ir Pt 57.1 e 1 Mn Fe Zn Ir Pt 48.4 b 5 Fe Co Ni Ir Pt 51.3 e 2 Mn Cu Zn Ir Pt 52.3 b 6 Fe Co Pd Ir Pt 49.6 e 3 Ni Zn Ru Pd Pt 39.6 b 7 Fe Co Ru Ir Pt 72.6 e 4 Co Ni Zn Ir Pt 49.4 b 8 Fe Co Zn Ir Pt 52.3 e 5 Mn Co Ni Ir Pt 49.5 b 9 Mn Fe Co Pd Pt 51.8 e 6 Mn Fe Co Ni Pt 47.3 b 10 Mn Fe Co Zn Pt 47.8 e 7 Mn Fe Cu Pd Pt 45.7 b 11 Fe Co Ni Pd Pt 58.6 e 8 Mn Fe Pd Ir Pt 44.4 b 12 Fe Co Ni Ru Pt 93.8 e 9 Mn Fe Zn Pd Pt 66.7 b 13 Fe Co Ni Zn Pt 54.2 e 10 Mn Cu Pd Ir Pt 44.4 c 0 Fe Co Zn Pd Pt 60.9 e 11 Mn Cu Zn Pd Pt 43.7 c 1 Fe Co Zn Ru Pt 63.4 e 12 Co Ni Pd Ir Pt 51.6 c 2 Mn Co Pd Ir Pt 64.7 e 13 Co Ni Zn Pd Pt 52.2 c 3 Mn Co Ru Ir Pt 49.1 f 0 Fe Cu Pd Ir Pt 42.3 c 4 Mn Co Zn Ir Pt 58 f 1 Fe Cu Zn Pd Pt 55.7 c 5 Co Ni Ru Ir Pt 34.7 f 2 Fe Zn Pd Ir Pt 45.5 c 6 Co Ru Pd Ir Pt 74.1 f 3 Mn Zn Pd Ir Pt 58.2 c 7 Co Zn Pd Ir Pt 82.5 f 4 Cu Zn Pd Ir Pt 47.6 c 8 Co Zn Ru Ir Pt 49 f 5 Mn Cu Ru Pd Pt 57.2 c 9 Mn Co Ni Pd Pt 69.1 f 6 Mn Fe Ru Pd Pt 47.5 c 10 Mn Co Ni Zn Pt 53.7 f 7 Mn Zn Ru Pd Pt 49.7 f 8 Mn Co Ru Pd Pt 67.5 h 4 Co Ni Ru Pd Pt 49.4 f 9 Mn Co Zn Pd Pt 50.4 h 5 Fe Co Ru Pd Pt 59.6 f 10 Co Zn Ru Pd Pt 42.6 h 6 Fe Cu Ru Pd Pt 49.4 f 11 Fe Ni Cu Ir Pt 51.7 h 7 Fe Ni Zn Ir Pt 54.2 f 12 Mn Fe Ni Cu Pt 131.9 h 8 Mn Fe Ni Pd Pt 48 f 13 Fe Ni Cu Pd Pt 72.6 h 9 Fe Ru Pd Ir Pt 55.7 g 0 Fe Ni Cu Ru Pt 35.7 h 10 Fe Zn Ru Pd Pt 54.5 g 1 Fe Ni Cu Zn Pt 41.6 h 11 Cu Zn Ru Pd Pt 51.2 g 2 Mn Ni Cu Ir Pt 45.6 h 12 Fe Cu Zn Ir Pt 51.4 g 3 Ni Cu Pd Ir Pt 25.5 h 13 Mn Co Cu Ru Pt 50.6 g 4 Ni Cu Ru Ir Pt 48.6 i 0 Mn Co Zn Ru Pt 48.5 g 5 ​ 96.2 i 1 ​ 51 g 6 ​ 37.7 i 2 ​ 44.4 g 7 ​ 56.1 i 3 ​ 48.5 g 8 ​ 36 i 4 ​ 42.5 g 9 ​ 63.4 i 5 ​ 53.7 g 10 ​ 40.5 i 6 ​ 51.6 g 11 ​ 43.6 i 7 ​ 56.2 g 12 ​ 34.7 i 8 ​ 38.7 g 13 ​ 46.4 i 9 ​ 49.7 h 0 ​ 35.5 i 10 ​ 58.4 h 1 ​ 49.1 i 11 ​ 49.5 h 2 ​ 41.6 i 12 ​ 52.3 h 3 ​ 48.8 i 13 ​ 43.2

[0062] As can be seen from Table 1, the combination of five-element high-entropy alloys obtained by the LLM-GA collaborative optimization and hybrid-coded high-entropy alloy multi-element combination intelligent screening method of the present invention has the best performance, its overpotential is lower than that of other combinations, and it can be obtained after only four generations of iteration, which effectively reduces the research and development cycle of high-entropy alloys.

[0063] Example 2 Dynamic Knowledge Base Construction

[0064] We obtained 79,664 documents from Web of Science and screened out 1,225 documents with high relevance and citation counts greater than 10 times / year (e.g. ​ ).

[0065] In the knowledge extraction stage, the literature was fine-tuned with LoRA using the XTuner toolkit to extract triplets of elements, activity coefficients, and synergy factors. A graph database (Neo4j) was constructed using this information to store the association graph between elements and performance.

[0066] In the graph, each node represents an element or performance indicator, and the weight of the edge represents the synergy coefficient between elements, and the weight value ranges from 0 to 1.

[0067] In terms of feature coding, element frequency statistics show that Pt, Fe, Co and Ni are high-frequency components (see ​ ), thus generating a five-tuple activity feature vector (e.g., Pt=0.92, Ni=0.78, Ir=0.65, Pd=0.59, Cu=0.43).

[0068] The knowledge base will be updated regularly, incorporating new experimental data through an incremental update mechanism every week.

[0069] The incremental update rule adopts the weighted average method, and the formula is K_new = 0.85 * K_old + 0.15 * Σφ(x_i), where φ(x_i) represents the characteristic mapping function of the experimental data.

[0070] In the preferred solution, to improve the accuracy of the synergy coefficient calculation, we used a graph attention network (GAT) to analyze the interactions between elements. This method uses 8 attention heads and sets an attention weight threshold of 0.3 to ensure that the calculated synergy coefficient accurately reflects the relationship between elements.

[0071] Example 3 High-throughput synthesis and characterization

[0072] This example describes the specific process of synthesis and testing. In the ultrafast thermal shock synthesis stage, a precursor solution containing 0.1M metal salt and carbon black was first prepared and heated to 600°C within 300 milliseconds using pulsed Joule heating technology, with a heating rate of 2000K / s and a cooling rate of 600K / s (see ​ ).

[0073] This method ensures rapid alloy synthesis and effectively avoids phase segregation. X-ray diffraction (XRD) analysis of the synthesized sample shows no obvious phase segregation, while transmission electron microscopy (TEM) analysis reveals uniform particles of 5 nanometers in size.

[0074] During electrochemical testing, an automated spotting robot evenly coated the catalyst ink onto a titanium plate at a loading of 0.2 mg / cm². The custom electrolytic cell implemented forced convection circulation to ensure uniform electrolyte distribution.

[0075] Linear sweep voltammetry (LSV) test showed that the overpotential of the PtNiIrPdCu / C catalyst was 28 mV.

[0076] In the preferred scheme, samples with XRD crystallinity below 80% are automatically rejected, triggering resynthesis. In addition, the electronic structure of the alloy is characterized, and these features are input into the fitness prediction module of the genetic algorithm for further optimization.

[0077] Example 4 Genetic Algorithm Optimization Process

[0078] This example describes the implementation details of the genetic algorithm. In the initial population generation stage, we randomly generate 6 five-element combinations and record their overpotential distribution, ranging from 20mV to 160mV (see ​ ).

[0079] In the evolutionary operation phase, crossover and mutation operations are used to generate new candidate individuals. Specifically, uniform crossover (crossover probability is 0.8) is used for the binary segment, and simulated binary crossover (distribution index η = 2) is used for the real segment.

[0080] The mutation operation includes flipping the binary segment with a probability of 0.02 and polynomial mutation of the real segment (mutation index η_m = 20).

[0081] In the directed replacement strategy, the lowest 20% of individuals with the lowest fitness were eliminated every five generations, and the element combination recommended by LLM (e.g., PtNiIrPdRh) was injected.

[0082] Through this series of operations, the genetic algorithm takes an average of 3.91 generations to converge in 20 independent trials (e.g. ​ Compared with the traditional method, the number of iterations is reduced by 60%.

[0083] In the optimal solution, the top 15% of individuals in each generation are retained and directly enter the next generation, thus ensuring the transmission of elite individuals. In addition, when the prediction error exceeds 10%, the mutation rate will automatically increase by 0.1 to improve the breadth and exploratory nature of the search.

[0084] The above embodiments describe in detail the structure, features and effects of the present invention. The above are only preferred embodiments of the present invention. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the scope covered by the specification, should be within the scope of protection of the present invention.

Claims

1. A high entropy alloy multi-element combination screening method based on LLM-GA collaborative optimization and hybrid coding, characterized in that: The following steps are involved: Step 1: Count the number of occurrences of various metal elements in all high-entropy alloy related literature, sort them from most to least according to the number of occurrences, and take the top n metal elements as the high-frequency metal elements; Step 2: Based on the metal elements screened in step 1, a hybrid coding genetic algorithm is used to generate a candidate metal combination; the hybrid coding genetic algorithm includes a binary segment and a real segment; The binary segment is the element selection segment, and the real segment is the proportional allocation segment; Step 3: Based on the candidate metal combinations generated in step 2, synthesize candidate alloys and test their properties; Step 4: Use the candidate alloys as individuals to construct a population. Based on the performance test results of step 3, use the fitness function to rank the alloys in the current population by performance, and eliminate individuals ranked in the bottom [20%] of the performance ranking, where [20%] refers to the integer value after rounding off. The fitness function is designed as F = 0.6 catalytic activity + 0.3 stability - 0.1 * precious metal cost; Step 5: Generate new metal combinations through hybrid coding genetic algorithm; Step 6: Iterate steps 2 to 5 until the individual with the best performance is obtained, that is, the alloy with the best metal combination performance is obtained.

2. The method according to claim 1, characterized in that The chromosome structure of the hybrid coding genetic algorithm includes an element selection segment and a proportion distribution segment, wherein the binary segment Ei∈{0,1} is used to identify the selection status of the element, 1 is selected and 0 is not selected; the real number segment Pi∈[0.05,0.35] represents the atomic percentage of each element, and satisfies the condition ΣPi=1, and the sum of the proportions of all elements is 100%.

3. A high entropy alloy multi-element combination screening system based on LLM-GA collaborative optimization and hybrid coding, characterized in that: include: Intelligent Document Analysis Module: This module builds a dynamic knowledge base from multi-source documents using a large language model (LLM), constructs a document semantic network based on the Transformer architecture, and generates element activity feature vectors using a multi-head attention mechanism. Genetic algorithm optimization module: uses binary-real mixed coding chromosomes, including element selection segments and proportion distribution segments, and implements a segmented crossover mutation strategy; Experimental verification module: Integrates high-throughput synthesis equipment with performance testing units and establishes a data feedback channel in JSON-LD format; Dynamic knowledge management subsystem: Builds an element-performance association graph based on a graph database, supports multimodal association between literature feature vectors and experimental data, and implements incremental updates of the knowledge base through version control.

4. The system according to claim 3, characterized in that The chromosome encoding structure of the genetic algorithm optimization module is [E1, E2, ..., En | P1, P2, ..., Pn], where: Ei∈{0,1} represents the element selection state, Pi∈[0.05,0.35] is the atomic percentage and satisfies ΣPi=1.

5. The system according to claim 3, characterized in that: The working steps of the genetic algorithm optimization module include: (1) Crossover operation is performed in segments: uniform crossover operation is used for binary coding segments; simulated binary crossover operation is used for real coding segments; (2) Mutation operations are implemented in a hierarchical manner: bit-flip mutation is performed on the binary segment; polynomial mutation is performed on the real segment; (3) Population elimination mechanism: Candidate alloys are used as individuals to construct a population. Based on the performance test results, the metal combinations in the current population are ranked by performance according to the preset fitness function, and the individuals ranked in the bottom 20% are eliminated. The number of individuals to be eliminated is an integer value rounded off to 20% of the total number of the current population. The fitness function is designed to be F = 0.6 catalytic activity + 0.3 stability - 0.1 * precious metal cost; (4) Large language model recommendation injection: The element ratios, performance data, and real-time parameters of the experimental verification module of the remaining metal combinations are input into the large language model to generate new element combinations that are at least the same as the number of eliminated ones; (5) Mixed population iteration: The new elements recommended by the large language model are combined with the retained original population to form a new generation of population for iterative optimization.

6. The system according to claim 3, characterized in that The updating rules of the dynamic knowledge base are: K_new = αK_old + (1-α) * Σφ(x_i) / n Where α = 0.85 is the knowledge decay factor, and φ(·) is the feature mapping function based on GAT.

7. The system according to claim 3, characterized in that: The experimental verification module includes: Ultrafast heating device, used to achieve rapid temperature treatment of materials, including one of a high-temperature thermal shock device, a laser heating device or a microwave heating device, wherein the heating rate of the high-temperature thermal shock device is ≥100℃ / s; Structural analysis equipment for monitoring microstructural changes in materials, including one or more of an in-situ X-ray diffraction (XRD) unit, a Raman spectrometer, or an infrared spectrometer; Performance testing equipment used to evaluate the functional properties of materials, including one or more of an electrochemical workstation, a mechanical strength tester, or a thermal conductivity tester; The experimental verification module transmits real-time data back to the genetic algorithm optimization module in JSON-LD format, triggering dynamic adjustment of genetic algorithm parameters.

8. The system according to claim 3, characterized in that: The system has a visual monitoring interface: The main curve dynamically displays the evolution trajectory of the optimal fitness of the genetic algorithm population; The secondary curve shows the trend of Shannon diversity index in parallel; A three-dimensional scatter plot maps the relationship between the catalytic activity and stability of element combinations and the cost of precious metals; Interactive control panel allows real-time adjustment of elite retention ratio (5%-15%), crossover rate (0.7-0.9), and mutation rate (0.01-0.05); The interface collaboratively presents the optimization process through multi-dimensional visualization, realizing dynamic intervention of algorithm parameters.

Images