Alloy composition aided design method based on large language model and knowledge graph

Abstract

The application belongs to the field of computer-aided design and emerging software technology, and relates to an alloy component aided design method based on a large language model and a knowledge graph, comprising: constructing a heterogeneous knowledge graph by using a language model and eliminating conflict relation edges based on a phase diagram rule matrix; generating a mechanism vector and a numerical vector, solving a tangent leading matrix under a local Riemann tangent space, and constructing an orthogonal tangent space collaborative projection matrix; accessing a phase transition boundary penalty factor to construct a high impedance penalty boundary in a collaborative representation space; retrieving an optimal mechanism path according to a target vector and decoding to solve each element mass percentage parameter; calculating a topological coincidence degree coefficient of a forward and reverse relation set, and outputting recommended parameters and a forward derivation graph topological chain when the standard is met; and the application is used as a solving optimization operator in computer-aided design software, eliminates non-physical continuity deviation caused by smooth interpolation, improves high-dimensional sparse space search convergence stability, and realizes closed-loop verification conforming to phase stability constraints.

Description

Technical Field

[0001] This invention relates to an alloy composition-aided design method based on a large language model and knowledge graph, belonging to the fields of computer-aided design and emerging software technology. Background Technology

[0002] Currently, alloy composition-aided design software typically trains machine learning models based on discrete numerical data and establishes a correspondence between the mass percentage of alloy composition and physical performance indicators to solve for the range of candidate component values. As the dimensionality and complexity of alloy systems increase, existing discrete numerical tables are easily limited by data sparsity and insufficient sample size. At the same time, a large amount of knowledge about phase transformation dynamics, grain boundary pinning mechanisms, and synergistic and antagonistic effects between elements is still scattered in literature texts. Conventional Euclidean space numerical regression algorithms cannot directly analyze unstructured knowledge, nor can they use it together with numerical data for calculation. Therefore, existing alloy composition-aided design software still faces bottlenecks in data representation and fusion when dealing with high-dimensional complex alloy design problems.

[0003] In the process of parameter design and solution of complex materials, traditional data representation algorithms usually regard the vector space as a flat Euclidean space and use continuous and smooth interpolation to align different features. However, there are often obvious nonlinear relationships and critical abrupt changes between alloy composition and physical properties. When the amount of a certain element added crosses the critical melting line, the microscopic phase state may undergo a step change. If smooth interpolation is still used, it is easy to treat the originally discontinuous changes at the phase transformation critical point as continuous changes. As a result, when the software performs extrapolation search outside the known data range, it may obtain mathematically continuous results that do not conform to the actual physical laws, and then output invalid parameter recommendations that violate the phase stability law. Moreover, it is difficult to give a clear logical derivation path for the entire calculation process.

[0004] Therefore, how to improve the spatial representation and projection of high-dimensional heterogeneous features from the software algorithm level under the emerging software and information technology service system, eliminate abnormal calculation results that do not conform to the phase transition law caused by smooth interpolation, and establish closed-loop verification and logically traceable deduction paths that conform to physical laws, so as to improve the stability of computer-aided design software in the design modeling and parameter solving process, remains a technical problem that needs to be solved. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of this invention is as follows: A method for alloy composition-aided design based on a large language model and knowledge graph, comprising the following steps: Step S1: Use a natural language processing model to extract triples of elements, micro-phases and performance indicators from documents and patent texts, construct a heterogeneous knowledge graph, and call up a preset phase graph rule matrix to eliminate logical conflict edges in the heterogeneous knowledge graph. Step S2: A graph embedding algorithm is used to generate mechanism vectors for a heterogeneous knowledge graph. The input alloy composition and performance numerical table is mapped into numerical vectors. The mechanism vectors are mapped to the local Riemann tangent space to solve for the tangent dominance matrix. The numerical vectors are subjected to covariance orthogonal decomposition to obtain the principal component directions. An orthogonal tangent space cooperative projection matrix is ​​constructed to constrain the orthogonal projection components of the principal component directions and the tangent dominance matrix. A phase transition boundary penalty factor constrained by the phase diagram rule matrix is ​​added to the optimization loss function. When the numerical vectors approach the phase transition dissolution line, a higher-order penalty gradient is calculated to construct a high-impedance penalty boundary in the cooperative representation space. Step S3: The received target performance instruction is converted into a target vector in the cooperative representation space. The optimal mechanism path is searched in the cooperative representation space based on the graph search algorithm. The orthogonal inverse projection matrix on the optimal mechanism path is calculated. The mass percentage parameters of each element are decoded and solved. Step S4: Input the calculated mass percentage parameters of each element into the reverse prediction model to generate a reverse relationship set, extract the triple relationship set corresponding to the best mechanism path as the forward relationship set, calculate the topological overlap coefficient between the forward relationship set and the reverse relationship set, and when the topological overlap coefficient is not less than 0.85, output the recommended parameters of each element's mass percentage and the corresponding forward derivation graph topological chain.

[0006] Preferably, step S1, which uses a natural language processing model to extract triples of elements, micro-phases, and performance indicators from literature and patent texts, specifically includes: using a natural language processing model pre-tuned on a materials science corpus to scan the input unstructured text, employing a character-level attention mechanism to locate the physical dependencies between element entities and performance indicator entities, and extracting triples of elements, micro-phases, and performance indicators; and retrieving a preset phase diagram rule matrix to eliminate logically conflicting relationship edges in the heterogeneous knowledge graph, specifically including: retrieving a preset thermodynamic phase diagram rule matrix to scan the heterogeneous knowledge graph, and when a relationship edge that violates the topological rules of phase instability is identified in the graph, removing the relationship edge or resetting its logical weight to zero.

[0007] Preferably, in step S2, a graph embedding algorithm is used to generate the mechanism vector of the heterogeneous knowledge graph. Specifically, a translation transformation graph embedding algorithm is used to map the nodes and relationships in the heterogeneous knowledge graph to a continuous low-dimensional vector space to generate the mechanism vector. The input alloy composition and performance numerical table is mapped into a numerical vector. Specifically, the input alloy composition and performance numerical table is input into a multilayer perceptron and the numerical vector is extracted as a continuous numerical feature vector.

[0008] Preferably, step S2, which involves constructing the orthogonal tangent space cooperative projection matrix and incorporating a phase transition boundary penalty factor constrained by the phase diagram rule matrix into the optimization loss function, includes the following steps: Step S21, calculating the tangent dominance matrix of the numerical vector in the local Riemann tangent space, and simultaneously performing covariance orthogonal decomposition on the numerical vector to obtain the principal component directions; Step S22, constructing the orthogonal tangent space cooperative projection matrix based on the tangent dominance matrix and the principal component directions, constraining the minimization of the orthogonal projection components of the principal component directions with the tangent dominance matrix under the tangent space transformation; Step S23, incorporating a phase transition boundary penalty factor determined by the thermodynamic phase diagram rules into the joint loss function. When the numerical vector approaches the phase transition melting line, the phase transition boundary penalty factor exhibits a stepwise increase in value, amplifying the geometric distance of the critical region in the cooperative representation space and constructing a high-impedance penalty boundary; Step S24, updating the orthogonal tangent space cooperative projection matrix based on the joint loss function, and training converges to obtain the cooperative representation space.

[0009] Preferably, step S3, which searches for the optimal mechanism path in the collaborative representation space based on the graph search algorithm, specifically includes: retrieving optimized mechanism paths from the target vector to the candidate element node vectors in the collaborative representation space, and activating the element interaction damping tensor during the optimization process. When the element combination currently sampled in the search path is marked as having a phase instability antagonistic relationship in the heterogeneous knowledge graph, the element interaction damping tensor automatically reduces the weight integral of the path. After the search is completed, the path with the highest cumulative weight is obtained as the optimal mechanism path.

[0010] Preferably, in step S3, the orthogonal inverse projection matrix on the optimal mechanism path is calculated, and the quality percentage parameters of each element are decoded and solved. Specifically, this includes: for the set of node vectors on the optimal mechanism path, the orthogonal inverse projection matrix of the orthogonal tangent space cooperative projection matrix is ​​used to calculate the node vector decoding, and the quality percentage range of each element that satisfies the target performance constraint is solved; wherein, the orthogonal inverse projection matrix is ​​used to inversely transform the low-dimensional vectors in the cooperative representation space back to the Euclidean numerical feature space.

[0011] Preferably, when the topological overlap coefficient is less than 0.85, the process of converting the set of differences into a penalty vector and feeding it back to step S3 to search for a new path specifically includes: mapping the missing edges and redundant edges in the set of differences into high-dimensional penalty gradients, superimposing them on the path weight matrix in the collaborative representation space, updating the sampling probability distribution of the search space, and triggering iterative optimization until the topological overlap coefficient reaches 0.85 or above.

[0012] Preferably, step S4 outputs the recommended parameters for the mass percentage of each element and the corresponding forward derivation graph topology chain, specifically including: rendering and outputting a component recommendation table containing the recommended median mass percentage of each element on a graphical interface, and simultaneously visually drawing an interpretable decision topology graph chain containing target performance, micro-enhancement mechanisms, and key element regulation nodes.

[0013] Compared with the prior art, the beneficial effects of the present invention are: 1. In alloy composition-assisted design using large language models and knowledge graphs, material mechanism knowledge from literature is mapped together with alloy composition and performance values ​​to a collaborative characterization space. Phase diagram rules are used to restrict relationships and search directions that do not conform to the laws of phase stability. This preserves the abrupt change characteristics that originally existed at the phase transformation critical point, avoids smooth interpolation from incorrectly treating the phase transformation boundary as a continuous change, reduces the number of candidate results that do not conform to physical laws generated by the software in the unknown parameter region, and improves the convergence stability of the parameter search process and the credibility of the recommendation results.

[0014] 2. In the process of searching for the optimal mechanism path, this invention reduces the weight of the corresponding path based on the antagonistic relationship of phase instability between elements, and solves the value range of candidate components by using an orthogonal inverse projection matrix. This allows the software to actively avoid component combinations that are prone to causing phase instability during the multi-element combination design process, and completes the solution of candidate parameters without the need for manual item-by-item screening, thereby reducing the divergence risk in the process of searching for complex alloy components.

[0015] 3. This invention verifies the topological overlap between the forward derivation results and the reverse prediction results of the recommended components, and feeds back the difference between the two to the path search process. When the verification result does not meet the requirements, the software can continue to adjust the candidate parameters; when the verification result meets the requirements, it can simultaneously provide the derivation relationship between the recommended components and their corresponding properties, microstructures, strengthening mechanisms, and key elements. This allows for closed-loop verification of the physical self-consistency of the recommended results, avoiding the output of discrete values ​​without a derivation basis, and improving the traceability of the logical derivation and the credibility of the output results in the alloy composition auxiliary design process. Attached Figure Description

[0016] Figure 1 This is a flowchart of the processing steps of the alloy composition-assisted design method of the present invention; Figure 2 This is a graph showing the change in the vector alignment factor and the value of the loss function in this invention.

[0017] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0019] An alloy composition-aided design method based on large language models and knowledge graphs includes the following steps: Step S1: Use a natural language processing model to extract triples of elements, micro-phases and performance indicators from documents and patent texts, construct a heterogeneous knowledge graph, and call up a preset phase graph rule matrix to eliminate logical conflict edges in the heterogeneous knowledge graph. Step S2: A graph embedding algorithm is used to generate mechanism vectors for a heterogeneous knowledge graph. The input alloy composition and performance numerical table is mapped into numerical vectors. The mechanism vectors are mapped to the local Riemann tangent space to solve for the tangent dominance matrix. The numerical vectors are subjected to covariance orthogonal decomposition to obtain the principal component directions. An orthogonal tangent space cooperative projection matrix is ​​constructed to constrain the orthogonal projection components of the principal component directions and the tangent dominance matrix. A phase transition boundary penalty factor constrained by the phase diagram rule matrix is ​​added to the optimization loss function. When the numerical vectors approach the phase transition dissolution line, a higher-order penalty gradient is calculated to construct a high-impedance penalty boundary in the cooperative representation space. Step S3: The received target performance instruction is converted into a target vector in the cooperative representation space. The optimal mechanism path is searched in the cooperative representation space based on the graph search algorithm. The orthogonal inverse projection matrix on the optimal mechanism path is calculated. The mass percentage parameters of each element are decoded and solved. Step S4: Input the calculated mass percentage parameters of each element into the reverse prediction model to generate a reverse relationship set, extract the triple relationship set corresponding to the best mechanism path as the forward relationship set, calculate the topological overlap coefficient between the forward relationship set and the reverse relationship set, and when the topological overlap coefficient is not less than 0.85, output the recommended parameters of each element's mass percentage and the corresponding forward derivation graph topological chain.

[0020] Preferably, step S1, which uses a natural language processing model to extract triples of elements, micro-phases, and performance indicators from literature and patent texts, specifically includes: using a natural language processing model pre-tuned on a materials science corpus to scan the input unstructured text, employing a character-level attention mechanism to locate the physical dependencies between element entities and performance indicator entities, and extracting triples of elements, micro-phases, and performance indicators; and retrieving a preset phase diagram rule matrix to eliminate logically conflicting relationship edges in the heterogeneous knowledge graph, specifically including: retrieving a preset thermodynamic phase diagram rule matrix to scan the heterogeneous knowledge graph, and when a relationship edge that violates the topological rules of phase instability is identified in the graph, removing the relationship edge or resetting its logical weight to zero.

[0021] Preferably, in step S2, a graph embedding algorithm is used to generate the mechanism vector of the heterogeneous knowledge graph. Specifically, a translation transformation graph embedding algorithm is used to map the nodes and relationships in the heterogeneous knowledge graph to a continuous low-dimensional vector space to generate the mechanism vector. The input alloy composition and performance numerical table is mapped into a numerical vector. Specifically, the input alloy composition and performance numerical table is input into a multilayer perceptron and the numerical vector is extracted as a continuous numerical feature vector.

[0022] Preferably, step S2, which involves constructing the orthogonal tangent space cooperative projection matrix and incorporating a phase transition boundary penalty factor constrained by the phase diagram rule matrix into the optimization loss function, includes the following steps: Step S21, calculating the tangent dominance matrix of the numerical vector in the local Riemann tangent space, and simultaneously performing covariance orthogonal decomposition on the numerical vector to obtain the principal component directions; Step S22, constructing the orthogonal tangent space cooperative projection matrix based on the tangent dominance matrix and the principal component directions, constraining the minimization of the orthogonal projection components of the principal component directions with the tangent dominance matrix under the tangent space transformation; Step S23, incorporating a phase transition boundary penalty factor determined by the thermodynamic phase diagram rules into the joint loss function. When the numerical vector approaches the phase transition melting line, the phase transition boundary penalty factor exhibits a stepwise increase in value, amplifying the geometric distance of the critical region in the cooperative representation space and constructing a high-impedance penalty boundary; Step S24, updating the orthogonal tangent space cooperative projection matrix based on the joint loss function, and training converges to obtain the cooperative representation space.

[0023] Preferably, step S3, which searches for the optimal mechanism path in the collaborative representation space based on the graph search algorithm, specifically includes: retrieving optimized mechanism paths from the target vector to the candidate element node vectors in the collaborative representation space, and activating the element interaction damping tensor during the optimization process. When the element combination currently sampled in the search path is marked as having a phase instability antagonistic relationship in the heterogeneous knowledge graph, the element interaction damping tensor automatically reduces the weight integral of the path. After the search is completed, the path with the highest cumulative weight is obtained as the optimal mechanism path.

[0024] Preferably, in step S3, the orthogonal inverse projection matrix on the optimal mechanism path is calculated, and the quality percentage parameters of each element are decoded and solved. Specifically, this includes: for the set of node vectors on the optimal mechanism path, the orthogonal inverse projection matrix of the orthogonal tangent space cooperative projection matrix is ​​used to calculate the node vector decoding, and the quality percentage range of each element that satisfies the target performance constraint is solved; wherein, the orthogonal inverse projection matrix is ​​used to inversely transform the low-dimensional vectors in the cooperative representation space back to the Euclidean numerical feature space.

[0025] Preferably, when the topological overlap coefficient is less than 0.85, the process of converting the set of differences into a penalty vector and feeding it back to step S3 to search for a new path specifically includes: mapping the missing edges and redundant edges in the set of differences into high-dimensional penalty gradients, superimposing them on the path weight matrix in the collaborative representation space, updating the sampling probability distribution of the search space, and triggering iterative optimization until the topological overlap coefficient reaches 0.85 or above.

[0026] Preferably, step S4 outputs the recommended parameters for the mass percentage of each element and the corresponding forward derivation graph topology chain, specifically including: rendering and outputting a component recommendation table containing the recommended median mass percentage of each element on a graphical interface, and simultaneously visually drawing an interpretable decision topology graph chain containing target performance, micro-enhancement mechanisms, and key element regulation nodes.

[0027] Example 1: This example deploys an alloy composition-aided design method based on a large language model and knowledge graph in a verification environment consisting of a computing server and a physical manufacturing end. The computing server uses a natural language processing model to extract triples of elements, microscopic phases, and performance indicators from alloy literature and patent texts, constructs a heterogeneous knowledge graph, and calls the thermodynamic phase diagram rule matrix to handle logically conflicting relationship edges. Subsequently, the mechanism vector of the heterogeneous knowledge graph and the numerical vector formed by alloy composition and performance data are mapped to a collaborative characterization space. The optimal mechanism path is retrieved according to the target performance instruction, and the mass percentage parameters of each element are decoded. The computing server generates a set of reverse relationships through a reverse prediction model, takes the set of triple relationships corresponding to the optimal mechanism path as the set of forward relationships, and calculates the topological overlap coefficient between the two. When the coefficient reaches 0.85, the recommended parameters of mass percentage of each element and the topological chain of the forward derivation graph are output. The physical manufacturing end receives the above output, converts the processing constraints corresponding to the recommended parameters into process parameters such as preheating temperature and heating time, and collects real-time physical response data during the manufacturing process, returning it to the computing server to calibrate the collaborative characterization space.

[0028] This embodiment uses the molding process of soft interior dashboard leather as the object of real-time physical response data collection. The leather has a large aspect ratio and a local variable cross-section topology. During the molding process, there are spatial nonlinear heat capacity heterogeneity and multi-axis alternating rotation field interference. The distributed adaptive control network includes a central scheduling unit and 12 independent heating zones arranged around the outside of the slush mold. Each independent heating zone is equipped with a local controller, temperature sensor and heating element.

[0029] During the cold preparation stage, the multi-axis drive mechanism drives the unloaded slush mold to perform a complete biaxial alternating rotation cycle. The central scheduling unit collects the stator current of the drive motor in the multi-axis drive mechanism at a sampling interval of 0.01s to establish an unloaded current reference time sequence, denoted as . ,in, Indicates the reference value of the no-load current. This indicates the rotation time index. Simultaneously, each local controller acquires the initial impedance reference of the corresponding heating element, denoted as... ,in, The initial impedance value is represented by the collected data, which is used to characterize the background noise generated by the inherent friction of the mechanical transmission chain, assembly clearance, and alternating gravitational torque, and serves as a reference for deducting background noise during the hot operation phase.

[0030] During the hot molding operation phase, powdered polyvinyl chloride (PVC) material is injected into the inner cavity of the slush mold. The multi-axis drive mechanism drives the slush mold to perform multi-axis alternating rotation according to a preset rotation trajectory. Simultaneously, the heating element is activated, and the central dispatch unit collects the stator current of the drive motor in real time, denoted as... ,in, This indicates the real-time stator current value. This indicates the rotation time index, and will display the current rotation time index. Corresponding real-time stator current value Subtract the no-load current reference value The net load current is obtained by subtracting the background noise of the mechanical transmission chain from the stator current signal. ,in, This represents the net load current value. The fluctuation is used to characterize the transient mass distribution and flow state of powdered polyvinyl chloride material in various local areas of the enamel mold cavity.

[0031] To reduce communication lag caused by global multidimensional matrix inversion, each local controller establishes a local adaptive optimization loop and an edge interlocking mechanism to implement dynamic thermal field control for each heating zone. The local controller obtains the real-time temperature value of the current heating zone through temperature sensors and calculates the temperature deviation between the target temperature value and the real-time temperature value, denoted as . ,in, Indicates the temperature deviation value, subscript This indicates the index of the heating zone; the local controller uses the temperature deviation value... and net load current value Adjust the pulse width modulation duty cycle of the current heating zone, denoted as . ,in, Indicates the value of pulse width modulation duty cycle, subscript This indicates the index of the heating zone.

[0032] Edge interlocking mechanisms have a duty cycle upper limit threshold. ,in, This indicates the maximum allowable pulse width modulation duty cycle threshold, which is set to 0.85 in this embodiment. When the pulse width modulation duty cycle of a specific heating zone... Exceeding the duty cycle upper limit threshold At that time, the sequence number index The corresponding local controller, through a distributed adaptive control network, sends signals to the sequence number indexed as follows: and The adjacent local controller sends an interlock suppression signal. Upon receiving this signal, the adjacent local controller reduces its current pulse width modulation duty cycle. The adjusted adjacent duty cycle is denoted as . And satisfy the following formula: ,in, This indicates the adjusted pulse width modulation duty cycle value for the adjacent heating zone. This indicates the original value of the pulse width modulation duty cycle of the adjacent heating zone before adjustment. This represents the cross-partition thermal coupling suppression coefficient, in this embodiment... The duty cycle is set to 0.45, and each local controller adjusts its duty cycle according to the above relationship to suppress thermal coupling between adjacent heating zones.

[0033] The finished leather material is prepared according to the following process: the weighed powdered polyvinyl chloride material is put into the material silo, the distributed adaptive control network is started, and the slush mold is rotated unloaded for 180s in a cold environment to complete data acquisition and benchmark establishment. When the initial temperature of each heating zone reaches the preset mold preheating temperature of 160℃, the slush mold is locked and connected with the material silo, and 5.4kg of powdered polyvinyl chloride material is injected into the mold cavity.

[0034] The multi-axis alternating rotation is initiated, with the main shaft rotation speed set to 10 r / min and the secondary shaft rotation speed set to 6 r / min. Simultaneously, the local adaptive optimization loop and edge interlocking mechanism are activated to adjust the pulse width modulation duty cycle in real time, maintain the thermal field balance of each heating zone, and continuously heat and mold for 120 seconds, so that the polymer powder melts on the inner wall of the mold and forms a uniform gel layer. After the material hopper recovers the excess material, the mold enters the cooling station, where the mold temperature is reduced to 50°C using a combination of air cooling and water mist cooling. Then, the robot arm grabs and demolds the mold to obtain the finished product of the automotive soft interior dashboard leather material.

[0035] According to the precision thickness gauge, the overall average wall thickness of the finished leather product is 1.2mm, and the wall thickness uniformity deviation at each measuring point is controlled within ±0.12mm. The local thinning or scorching defects caused by the variable cross-section topology have been eliminated, and the product defect rate is less than 0.3%.

[0036] During the verification process at the physical manufacturing stage, the distributed adaptive control network converts the material processing constraints corresponding to the recommended parameters of each element's mass percentage into process parameters such as a preheating temperature of 160℃ and a heating time of 120s, and also converts the net load current... Temperature deviation The real-time physical response data is returned to the computing server to calibrate the collaborative representation space formed by the natural language processing model and the heterogeneous knowledge graph, thus realizing a closed loop from alloy composition parameter calculation and process parameter execution to physical response data feedback.

[0037] Example 2: The current test system is deployed in a computing server containing a central processing unit and a graphics processing unit. The central processing unit has an instruction processing throughput of 120 TFLOPS, a volatile memory capacity of 128 GB, and a network interface data transmission bandwidth of 10 Gbps. The test data includes publicly available materials science text datasets MatSci-Corpus and experimental records of high-temperature nickel-based alloys collected by the physical experimental platform. The physical experimental platform is equipped with a scanning electron microscope with a spatial resolution of 1 nm and an X-ray energy dispersive spectrometer with an elemental detection sensitivity of 0.1 wt%.

[0038] The natural language processing model used in the test was a large language model pre-tuned on a materials domain corpus. The central processing unit read the high-temperature alloy text data. During multi-task parallel processing, the allocator distributed global tasks according to the proportion of the current free storage capacity of each computing unit in the cluster. The natural language processing model scanned the allocated unstructured text and located element entities, micro-phase entities, and performance index entities through a character-level attention mechanism. First, a coarse extraction entity set containing the names and mass fractions of the formulation components, the solution aging temperature, and the holding time was formed. Then, triplets of elements, micro-phases, and performance indexes were extracted according to the physical dependencies of each entity in the text.

[0039] The extracted triples are used to construct a heterogeneous knowledge graph. The thermodynamic phase diagram rule matrix scans the relation edges corresponding to each triple. When a relation edge violates the phase instability topology rule, the relation edge is removed from the heterogeneous knowledge graph or its logical weight is reset to zero. The heterogeneous knowledge graph has a periodic reconstruction mechanism. Whenever 1000 new alloy experimental records are imported, the system starts graph re-verification and recalibrates the connection relationships between nodes.

[0040] To test the performance of the natural language processing model and heterogeneous knowledge graph on irregular engineering corpora, 15% textual grammatical noise and 10% element symbol annotation errors were injected into the original text data. The entity retrieval step threshold for the heterogeneous knowledge graph was denoted as [missing information]. This threshold is used to limit the maximum number of hops for graph topology traversal. It is determined based on the node connection density of the heterogeneous knowledge graph and the computation cycle limitations of the central processing unit, while also considering the path coverage of alloy elements and the memory allocation load of the computing server. When the average node degree is between 5 and 15, to ensure that the graph path coverage is not less than 90%, [the threshold is set to -1]. The value is limited to between 2 and 4, and is selected according to an integer relationship that is inversely proportional to the average node degree; under the operating condition where the average node degree is 8, Take 3.

[0041] In the structured entity alignment process, the text candidate triples output by the natural language processing model are matched item by item with the nodes and relations in the heterogeneous knowledge graph, and the alignment confidence score is denoted as... And calculate using the following formula: ,in, This indicates the number of alloy element triples that were successfully matched in the heterogeneous knowledge graph. This represents the total number of candidate triples in the text output by the natural language processing model.

[0042] In the control group, Method 1, the threshold for the number of entity retrieval steps. The number of successfully matched alloy element triples is 0. The total number of candidate triples in the text is 38. The value is 72. Substituting the two data points into the above formula, the resulting alignment confidence score is... The value was 0.528, the alloy strength prediction error was 14%, and the calculation time was 110ms. In the control group, Method 2, Also equal to 0, It is 41. It is 70, corresponding to The value was 0.586, the alloy strength prediction error was 12%, and the calculation time was 105ms. Neither of the two control groups utilized the retrieval step expansion graph topology association path. Under input conditions containing text syntax noise and element symbol labeling errors, the strength prediction error was not less than 12%.

[0043] Using the method of this embodiment and When set to 1, the heterogeneous knowledge graph only searches within a small number of hops, and has not yet fully established a deep evolutionary relationship between formulation components, proportions, heat treatment process parameters, and alloy phase structure. Under this condition, It is 48. The calculated value is 71. The value is 0.676, the alloy strength prediction error is 9%, and the calculation time is 145ms.

[0044] Will When set to 3, the heterogeneous knowledge graph constrains entities and relationships in noisy text through topological associations. It is 69. The corresponding value is 72. The value is 0.958. At this point, the prediction error for the high-temperature alloy strength is 3%, and the calculation time is 220ms. Continuing... When increased to 5, It is 70. It is 73. The value is 0.959, and the intensity prediction error remains at 3%. Due to the expanded graph topology search range, the computation time increases to 890ms. As can be seen from the above data, in After setting the value to 3, the alignment confidence score and intensity prediction error have stabilized. Further expanding the search scope did not decrease the intensity prediction error further. Therefore, the following approach was adopted. The output of the high-temperature alloy composition prediction results is for operating condition 3.

[0045] During the forward and reverse verification process, the calculated quality percentage parameters of each element are input into the reverse prediction model. The reverse prediction model generates triples according to the node identifiers and relation types in the heterogeneous knowledge graph, forming a reverse relation set. At the same time, entity and relation triples are extracted from the optimal mechanism path to form a forward relation set. When triples in two sets have the same node identifier and relation type, they are counted as overlapping elements. The topological overlap coefficient is calculated by dividing the number of intersection elements of the forward relation set and the reverse relation set by the number of union elements of the two sets. When the two sets have no intersection, the coefficient is 0, and when the two sets completely overlap, it is 1.00.

[0046] The threshold for determining the topological overlap coefficient was determined by comparing 1000 sets of standard high-entropy alloy experimental samples. When the coefficient is below 0.85, the proportion of mechanism breaks in the forward and reverse topological chains is at least 15%, and the prediction error increases to more than 9%. When the coefficient reaches or exceeds 0.85, the closed-loop verification accuracy of the physical mechanism spectrum reaches more than 95%, and the alloy strength prediction error is stable within 3%. Therefore, 0.85 is taken as the critical lower limit for the set of forward and reverse relationships to meet the closed-loop verification requirements.

[0047] When the topological overlap coefficient is below 0.85, the system compares the positive relation set and the negative relation set, extracts missing edges and redundant edges to form a difference relation set, and maps the nodes and relations in the difference relation set to high-dimensional penalty gradients. These high-dimensional penalty gradients are superimposed on the path weight matrix of the collaborative representation space to reduce the sampling probability of the corresponding path, and the optimal mechanism path search is performed again. When the topological overlap coefficient reaches 0.85 or above, the iteration stops and the recommended parameters for the quality percentage of each element and the topological chain of the forward derivation graph are output.

[0048] When graphics processor resources are limited, the computing server switches to a matching mode between the bidirectional long short-term memory network and the rule graph. After adopting this mode, the strength prediction error increases by 2% and the computation time decreases by 50%. When the computing resources meet the requirements of heterogeneous knowledge graph retrieval, the natural language processing model and the heterogeneous knowledge graph convert the alloy design text data containing multiple noises into a structured vector constrained by the graph topology. The system outputs an alloy high-temperature tensile strength prediction error of 3%.

[0049] Example 3: This example focuses on the composition optimization of local microstructure phase transformation and high-temperature tensile properties of high-entropy titanium alloys. High-temperature conditions have a synergistic constraint on the yield strength and elongation of the material. The central processing unit in the computing platform screens candidate components based on multi-source metallurgical texts and heterogeneous knowledge graphs. The central processing unit reads unstructured alloy literature semantic texts from memory, as well as a topological structure matrix containing titanium alloy phase transformation node entities. These two are used as the first technical object feature and the second technical object feature at the input end, respectively.

[0050] The natural language processing model uses a large language model to generate labeled sequences of candidate alloy components and extracts semantic vectors of candidate titanium alloy components. The heterogeneous knowledge graph uses a translation transformation graph embedding algorithm to map nodes and relationships to a continuous low-dimensional vector space to generate the mechanism vector of the target alloy. The mass percentage of each element and the corresponding performance values ​​of the candidate components are input into a multilayer perceptron to extract numerical vectors and perform candidate component screening in a collaborative representation space.

[0051] During the process of the natural language processing model outputting candidate alloy components, the semantic vector and mechanism vector of the candidate titanium alloy components may deviate at the local mapping position, causing the overall proportion of the candidate components to deviate from the constraint range of the thermodynamically stable phase, and resulting in the failure of subsequent virtual simulation to converge. In order to identify this deviation, the central processing unit calls the attention mechanism layer of the natural language processing model and calculates the spatial alignment deviation value.

[0052] The central processing unit (CPU) maps the semantic vectors of candidate titanium alloy compositions and the mechanism vectors of the target alloy, both labeled with metallurgical semantic tags, to a shared Euclidean space. It then calculates the norm 2 modulus of both, using the norm 2 modulus of the candidate titanium alloy composition semantic vector as the first eigenvalue and the norm 2 modulus of the mechanism vector as the corresponding reference eigenvalue. Both are scalars of the same dimension. The CPU calculates the space alignment deviation using the following formula: ,in, Indicates the spatial alignment deviation value, subscript Representing spatial topology; The subscript represents the first eigenvalue of the semantic vector of candidate titanium alloy composition in the shared Euclidean space. Represents semantic features; The subscript represents the corresponding reference eigenvalue of the target alloy mechanism vector in the shared Euclidean space. Represents a topological entity.

[0053] The first eigenvalue of the semantic vector of candidate titanium alloy composition within one computation cycle The value is 0.843, which is the corresponding reference eigenvalue of the target alloy mechanism vector. The value is 0.716. The central processing unit calculates the difference between the two to obtain the spatial alignment deviation value. The value is 0.127. The central processing unit sets a sliding time window containing 50 historical titanium alloy experimental records in the memory. Each time new experimental data is input, the sliding time window discards the oldest set of data and loads the latest data. The central processing unit reads the spatial alignment deviation value of the 50 sets of historical data in the window, calculates the standard deviation of the historical alignment error, and then multiplies the standard deviation by 2 to obtain the spatial alignment deviation threshold for the current cycle. In this running cycle, the standard deviation corresponding to the 50 sets of historical data is 0.048, and the spatial alignment deviation threshold is 0.096.

[0054] The central processing unit (CPU) compares the current spatial alignment deviation value of 0.127 with the spatial alignment deviation threshold of 0.096. Since 0.127 is greater than 0.096, the current candidate titanium alloy composition exceeds the constraint range of the thermodynamically stable phase. The CPU interrupts the output of this tag sequence and uses the spatial alignment deviation value as a negative feedback error, which is then passed back to the attention mechanism weight matrix of the natural language processing model to adjust the output of the next round of candidate components. After adjustment, the spatial alignment deviation value of the new candidate component is reduced to 0.058. Since this value is not greater than 0.096, the CPU outputs the candidate component to the downstream simulation bus.

[0055] After the candidate components enter the collaborative characterization space, the central processing unit (CPU) calls the thermodynamic phase diagram rule matrix to extract the phase transformation dissolution line parameters of the corresponding alloy system. When the distance between the numerical vector and the phase transformation dissolution line is less than 0.02, the phase transformation boundary penalty factor based on the exponential function is activated, and it increases quadratically with decreasing distance. The CPU uses the value of the phase transformation boundary penalty factor as a weighting term in the joint loss function to obtain the higher-order penalty gradient of the numerical vector. When the numerical vector continues to approach the phase transformation dissolution line, the gradient direction is opposite to the direction of its continued entry into the unstable phase transformation region, thus forming a high-resistance penalty boundary in the collaborative characterization space.

[0056] When conducting a control experiment using the method of this embodiment, the processor's basic computing power is [value missing] per second. The original historical experimental dataset collected by the sensor contains 10,254 records of titanium alloy composition ratios. The sampling frequency of the graph embedding algorithm for nodes and relationships in the heterogeneous knowledge graph is set to 50Hz. During a continuous 512-hour period, the spatial alignment deviation value in the test group using the method of this embodiment remains below the spatial alignment deviation threshold.

[0057] The control experiment also included a partially missing control group. This control group kept the parameters of the natural language processing model unchanged and removed the feedback loop of spatial alignment deviation value based on the heterogeneous knowledge graph topology matrix. When calculating the candidate components of high-entropy titanium alloy, the proportion of invalid components in the partially missing control group was 42.3%. In the test group using the method of this embodiment, the proportion of invalid components was less than 3.8%, and the convergence speed of candidate components was improved by 35.6%.

[0058] Example 4: This example combines Figures 1 to 2 This paper describes an alloy composition-aided design method based on large language models and knowledge graphs, such as... Figure 1 As shown, the steps S1, S2, S3, and S4 are connected sequentially from top to bottom by arrows. Step S1 uses a natural language processing model to extract triples of elements, microscopic phases, and performance indicators from documents and patent texts to construct a heterogeneous knowledge graph. A preset phase diagram rule matrix is ​​then retrieved to eliminate logically conflicting edges in the heterogeneous knowledge graph. Step S2 uses a graph embedding algorithm to generate the mechanism vector of the heterogeneous knowledge graph. The input alloy composition and performance numerical tables are mapped into numerical vectors. The mechanism vectors are mapped to a local Riemann tangent space to solve for the tangent dominance matrix. Covariance orthogonal decomposition is performed on the numerical vectors to obtain the principal component directions. An orthogonal tangent space co-projection matrix is ​​constructed, and the projected components are constrained to incorporate phase transitions in the optimized loss function. Boundary penalty factor: When approaching the phase transition melting boundary, a high impedance penalty boundary is constructed. Step S3: The received target performance instruction is transformed into a target vector in the cooperative representation space. Based on the graph search algorithm, the optimal mechanism path is searched in the cooperative representation space, and the orthogonal inverse projection matrix on the optimal mechanism path is calculated. The mass percentage parameters of each element are decoded and solved. Step S4: The calculated mass percentage parameters of each element are input into the inverse prediction model to generate an inverse relation set. The triple relation set corresponding to the optimal mechanism path is extracted as the forward relation set. The topological overlap coefficient between the forward relation set and the inverse relation set is calculated. When the topological overlap coefficient is not less than 0.85, the recommended parameters of the mass percentage of each element and the corresponding forward derivation graph topological chain are output.

[0059] like Figure 2The figure illustrates the response mapping relationship between the vector alignment factor β and the loss function value L. A legend is displayed above the figure, where solid lines represent the joint loss function values, and dashed lines represent the loss convergence tolerance threshold. The horizontal dimension represents the vector alignment factor β, with values ​​labeled 0.1, 0.2, 0.35, 0.5, 0.65, 0.8, and 0.9. The vertical dimension represents the loss function value L, with values ​​labeled 0, 0.01, 0.02, 0.03, 0.04, and 0.05. The solid lines represent... The value of the joint loss function exhibits a U-shaped curve trend of first decreasing and then increasing with the increase of the vector alignment factor β. It decreases before the vector alignment factor β is between 0.1 and 0.5 and extends to about 0.60. It reaches the lowest point below the loss convergence tolerance threshold shown by the dashed line when the vector alignment factor β is about 0.60. Then the curve turns to an increasing trend. When the vector alignment factor β reaches 0.65 and subsequently increases to 0.9, the value of the joint loss function continues to rise and is again above the loss convergence tolerance threshold.

[0060] Example 5: This example uses a hardware processor system to perform alloy composition-aided design. The system includes an input interface, a knowledge graph retrieval unit, a large language model prediction unit, and a physical constraint filtering loop. The input interface collects the mechanical property design indicators of the target alloy, including the target tensile strength and target elongation values, and converts them into target vectors. The knowledge graph retrieval unit receives the target vector. The system calls upon a pre-defined ternary phase diagram database for aluminum-based alloys to retrieve the initial composition range of alloying elements corresponding to the target tensile strength value. .

[0061] The large language model prediction unit runs a natural language processing model and converts the target vector... and initial component range Convert to a sequence of text tags The self-attention mechanism layer computes the text tag sequence. The interaction weights between each marker are used to output the probability distribution of alloy composition prediction. To filter outlier data in the probability distribution, a physical constraint filtering loop is connected to the output of the attention mechanism layer. The allocation ratios of the top three candidate alloy groups are extracted from this probability distribution, and the lattice mismatch parameters are calculated for each group. The calculation relationship is as follows: ,in, Indicates the degree of misfit in the alloy lattice. Indicates the first The atomic percentage of each alloying element Indicates the first The atomic radius of each alloying element, Indicates the average atomic radius of the mixed system; subscript Indicates mismatch, subscript Indicates element index, subscript This represents the average value.

[0062] The physical constraint filtering loop assigns the corresponding lattice mismatch parameters to each candidate alloy group. Compared to the stable solid solution threshold of 0.05, when When the value is greater than 0.05, it is determined that the corresponding candidate alloy group distribution ratio has a risk of precipitating harmful phases. A negative feedback signal is sent to the input of the large language model prediction unit to adjust the bias parameters of the self-attention mechanism layer and re-output the alloy composition prediction probability distribution. When the value is less than or equal to 0.05, the candidate alloy group is determined to meet the constraints of a stable single-phase solid solution structure, and the recommended parameters of the corresponding mass percentage of each element are output as the alloy composition design scheme.

[0063] When searching for the optimal mechanism path in the collaborative representation space, the element interaction damping tensor adopts a three-dimensional symmetric matrix, whose three dimensions correspond to the first element, the second element, and the micro-phase state to be combined, respectively. When the element combination sampled by the current search path has a phase instability antagonistic relationship in the heterogeneous knowledge graph, the damping coefficient corresponding to the element combination and the micro-phase state is activated. The value of the damping coefficient is between 0.10 and 0.50. It is multiplied by the cumulative weight of the current path, and the result is deducted from the path weight integral as a deduction to reduce the cumulative weight of the corresponding path. After the path search is completed, the path with the highest cumulative weight is determined as the optimal mechanism path.

[0064] The proportions of each component in the alloy composition design are determined based on the material physics mechanism. Taking the mass percentage of titanium as an example, its value ranges from 1.5% to 3.5%. When the mass percentage of titanium is less than 1.5%, the number of aging precipitates is insufficient, the resistance to dislocation movement within the grains is weakened, and the tensile strength of the alloy is less than 600 MPa. When the mass percentage of titanium is greater than 3.5%, coarse eutectic structures accumulate at the alloy grain boundaries, inducing early microcrack initiation under the action of applied tensile stress, reducing the elongation of the alloy to below 3%. When the mass percentage of titanium is between 1.5% and 3.5%, titanium undergoes atomic-level synergistic coupling with aluminum and zirconium elements in the system. Under the aging treatment condition of 350℃, a dispersed phase with an average particle size of 15 nm precipitates inside the matrix. This dispersed phase hinders the climbing movement of dislocations within the grains, keeping the tensile strength of the alloy between 620 MPa and 650 MPa, while keeping the elongation above 8%.

[0065] Example 6: During the initial system deployment, the processing chip calibrates the text tag vectors output by the natural language processing model and the mechanism vectors of the heterogeneous knowledge graph in an offline state. This converts alloy crystal structure data and text semantic features from different sources into a unified-dimensional digital matrix. The initial mapping range is determined by the feature mapping dimension. Confirmed, in this embodiment Using a value of 512, the processing chip calculates the Euclidean distance between the node vectors of the heterogeneous knowledge graph and the text word vectors, and then uses a vector alignment factor. Adjust the weight ratio of the two types of features; adjust when the input alloy composition data fluctuates. To maintain the stability of geometric distances within the mapping space, a baseline mapping matrix is ​​formed for subsequent alloy composition prediction.

[0066] When constructing the orthogonal tangent space cooperative projection matrix, the tangent vectors of the mechanism vectors generated by the graph embedding algorithm are first obtained along multiple directions in the local neighborhood. The tangent vectors form a local basis matrix. Then, the eigenvectors corresponding to the first three largest eigenvalues ​​of this matrix are selected to form the tangent dominance matrix. For the continuous feature array formed by alloy composition and performance values, the processing chip performs covariance orthogonal decomposition and extracts the principal component directions with a cumulative variance contribution rate of 95%. The processing chip takes the tangent dominance matrix and the principal component directions as inputs and performs a Schmitt orthogonalization transformation on the principal component directions so that the orthogonal projection components of the principal component directions relative to the tangent dominance matrix under the tangent space transformation approach zero. Based on this, the orthogonal tangent space cooperative projection matrix is ​​solved. This matrix maps the mechanism vectors and numerical vectors to the same cooperative representation space, so that the mechanism directions in the non-Euclidean manifold are orthogonally aligned with the Euclidean numerical features.

[0067] To determine the vector alignment factor The system establishes a loss function. The calculation relationship is as follows: ,in, This represents the representation error of heterogeneous knowledge graphs. This represents the semantic alignment error of the natural language processing model. The system iterates and adjusts the error in increments of 0.05 within a range of 0.1 to 0.9. until the loss function Reaching the preset convergence state, with a loss of convergence tolerance. Set to 0.001, number of iterations. Set the value to 1000; when the iteration reaches... And loss function The change range is less than When the time comes, the optimization process ends.

[0068] In the loss function described above, after combining like terms, the vector alignment factor... The weights used to adjust the representation error of heterogeneous knowledge graphs and the semantic alignment error of natural language processing models, when When the value is below 0.1, the mapping result becomes more dependent on the semantic features of the text, and candidate components are more likely to deviate from the physical topological constraints in the heterogeneous knowledge graph; when When the value is higher than 0.9, the mapping results are mainly limited by existing spectral relationships, narrowing the prediction range for unknown alloy compositions. After completing the traversal with a step size of 0.05, When the value is 0.65, the loss function It reaches the minimum convergence point.

[0069] Within a defined collaborative characterization space, the system verifies the composition of nickel-based superalloys and sets the spectral node density threshold. Set to 0.75, the processing chip retrieves grain boundary strengthening and solid solution strengthening entities of nickel-based superalloys from the heterogeneous knowledge graph and maps them to tensile strength and elongation data in the literature. When the vector alignment factor is set to 0.75... When the value stabilizes at 0.65, the deviation between the system's recommended alloy composition and the actual smelting test data is less than 3.5%. During subsequent continuous operation, the system continues to use this mapping state to respond to external composition design instructions and output candidate compositions for nickel-based superalloys.

[0070] When inversely mapping the solution vectors in the co-representation space to the Euclidean numerical feature space, the system uses the orthogonal inverse projection matrix corresponding to the orthogonal tangent space co-projection matrix to perform matrix operations on the set of node vectors on the optimal mechanism path to obtain the component eigenvectors in the Euclidean numerical feature space. The system then reads the mass percentage value range according to the characteristic dimension corresponding to each element. At the inverse mapping interface, the system applies a piecewise step truncation operator according to the thermodynamic phase diagram rule matrix. When the inverse transformation trajectory approaches the phase transition dissolution line, the operator maps the high impedance penalty boundary in the co-representation space to a discontinuous point in the Euclidean numerical feature space, causing the inverse transformation trajectory to stop crossing the phase transition dissolution line. As a result, the mass percentage parameters of each element obtained retain the numerical expression in the Euclidean numerical feature space and avoid the unstable region of phase transition.

[0071] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for alloy composition-aided design based on large language models and knowledge graphs, characterized in that, Includes the following steps: Step S1: Use a natural language processing model to extract triples of elements, micro-phases and performance indicators from documents and patent texts, construct a heterogeneous knowledge graph, and call up a preset phase graph rule matrix to eliminate logical conflict edges in the heterogeneous knowledge graph. Step S2: A graph embedding algorithm is used to generate mechanism vectors for a heterogeneous knowledge graph. The input alloy composition and performance numerical table is mapped into numerical vectors. The mechanism vectors are mapped to the local Riemann tangent space to solve for the tangent dominance matrix. The numerical vectors are subjected to covariance orthogonal decomposition to obtain the principal component directions. An orthogonal tangent space cooperative projection matrix is ​​constructed to constrain the orthogonal projection components of the principal component directions and the tangent dominance matrix. A phase transition boundary penalty factor constrained by the phase diagram rule matrix is ​​added to the optimization loss function. When the numerical vector approaches the phase transition dissolution line, a higher-order penalty gradient is calculated to construct a high-impedance penalty boundary in the cooperative representation space. Step S3: The received target performance instruction is converted into a target vector in the cooperative representation space. The optimal mechanism path is searched in the cooperative representation space based on the graph search algorithm. The orthogonal inverse projection matrix on the optimal mechanism path is calculated. The mass percentage parameters of each element are decoded and solved. Step S4: Input the calculated mass percentage parameters of each element into the reverse prediction model to generate a reverse relationship set, extract the triple relationship set corresponding to the best mechanism path as the forward relationship set, calculate the topological overlap coefficient between the forward relationship set and the reverse relationship set, and when the topological overlap coefficient is not less than 0.85, output the recommended parameters of each element's mass percentage and the corresponding forward derivation graph topological chain.

2. The alloy composition-aided design method based on a large language model and knowledge graph according to claim 1, characterized in that, Step S1 involves using a natural language processing model to extract triples of elements, micro-phases, and performance indicators from literature and patent texts. Specifically, this includes: using a pre-tuned natural language processing model on a materials science corpus to scan the input unstructured text, employing a character-level attention mechanism to locate the physical dependencies between element entities and performance indicator entities, and extracting triples of elements, micro-phases, and performance indicators; and retrieving a preset phase diagram rule matrix to eliminate logically conflicting edges in the heterogeneous knowledge graph. Specifically, this includes: retrieving a preset thermodynamic phase diagram rule matrix to scan the heterogeneous knowledge graph, and when an edge that violates the topological rules of phase instability is identified in the graph, removing the edge or resetting its logical weight to zero.

3. The alloy composition-aided design method based on a large language model and knowledge graph according to claim 1, characterized in that, In step S2, a graph embedding algorithm is used to generate the mechanism vector of the heterogeneous knowledge graph. Specifically, the translation transformation graph embedding algorithm is used to map the nodes and relationships in the heterogeneous knowledge graph to a continuous low-dimensional vector space to generate the mechanism vector. The input alloy composition and performance numerical table is mapped into a numerical vector. Specifically, the input alloy composition and performance numerical table is input into a multilayer perceptron and the numerical vector is extracted as a continuous numerical feature vector.

4. The alloy composition-aided design method based on a large language model and knowledge graph according to claim 1, characterized in that, Step S2 involves constructing an orthogonal tangent space cooperative projection matrix and incorporating a phase transition boundary penalty factor constrained by a phase diagram rule matrix into the optimization loss function. This includes the following steps: Step S21: Calculate the tangent dominance matrix of the numerical vector in the local Riemann tangent space, and simultaneously perform covariance orthogonal decomposition on the numerical vector to obtain the principal component directions; Step S22: Construct an orthogonal tangent space cooperative projection matrix based on the tangent dominance matrix and the principal component directions, constraining the minimization of the orthogonal projection components of the principal component directions with the tangent dominance matrix under tangent space transformation; Step S23: Incorporate a phase transition boundary penalty factor determined by thermodynamic phase diagram rules into the joint loss function. When the numerical vector approaches the phase transition melting line, the phase transition boundary penalty factor exhibits a stepwise increase in value, amplifying the geometric distance of the critical region in the cooperative representation space and constructing a high-impedance penalty boundary; Step S24: Update the orthogonal tangent space cooperative projection matrix based on the joint loss function, and train until convergence to obtain the cooperative representation space.

5. The alloy composition-aided design method based on a large language model and knowledge graph according to claim 1, characterized in that, In step S3, the optimal mechanism path is searched in the collaborative representation space based on the graph search algorithm. Specifically, this includes: retrieving the optimized mechanism path connecting the target vector to the candidate element node vector in the collaborative representation space, and activating the element interaction damping tensor during the optimization process. When the element combination currently sampled in the search path is marked as having a phase instability antagonistic relationship in the heterogeneous knowledge graph, the element interaction damping tensor automatically reduces the weight integral of the path. After the search is completed, the path with the highest cumulative weight is obtained as the optimal mechanism path.

6. The alloy composition-aided design method based on a large language model and knowledge graph according to claim 1, characterized in that, In step S3, the orthogonal inverse projection matrix on the optimal mechanism path is calculated, and the quality percentage parameters of each element are decoded and solved. Specifically, this includes: for the set of node vectors on the optimal mechanism path, the orthogonal inverse projection matrix of the orthogonal tangent space cooperative projection matrix is ​​used to calculate the node vector decoding, and the range of quality percentage values ​​of each element that satisfies the target performance constraints is solved. Among them, the orthogonal inverse projection matrix is ​​used to inversely transform the low-dimensional vectors in the cooperative representation space back to the Euclidean numerical feature space.

7. The alloy composition-aided design method based on a large language model and knowledge graph according to claim 1, characterized in that, When the topological overlap coefficient is less than 0.85, the process of converting the set of differences into a penalty vector and feeding it back to step S3 to search for a new path includes: mapping the missing and redundant edges in the set of differences into high-dimensional penalty gradients, superimposing them on the path weight matrix in the collaborative representation space, updating the sampling probability distribution of the search space, and triggering iterative optimization until the topological overlap coefficient reaches 0.85 or above.

8. The alloy composition-aided design method based on a large language model and knowledge graph according to claim 1, characterized in that, Step S4 outputs the recommended parameters for the mass percentage of each element and the corresponding forward derivation graph topology chain. Specifically, it includes rendering and outputting a component recommendation table containing the median recommended mass percentage of each element on the graphical interface, and simultaneously visually drawing an interpretable decision topology graph chain containing target performance, micro-enhancement mechanisms, and key element regulation nodes.

Images