A deep learning-based intelligent optimization method for communication chip design parameters

Abstract

This invention discloses an intelligent optimization method for communication chip design parameters based on deep learning, specifically including: reading communication chip design data to generate a parameter vector sequence and a target performance vector sequence; encoding the parameter vector sequence to generate a parameter representation vector sequence; constructing a deep learning mapping model to generate a performance prediction vector sequence; performing difference calculation to generate an error vector sequence; constructing a parameter relationship structure, generating a random measure, performing time evolution, and calculating entropy flow constraints to regulate parameter updates; performing parameter updates based on the error vector sequence, random measure, and entropy flow constraints; calculating parameter influence to generate a density modulation function and performing weighted modulation on the parameters; and performing constraint determination to generate the optimization result of the communication chip design parameters. This invention achieves adaptive optimization and multi-constraint collaborative regulation of communication chip design parameters, reduces parameter convergence errors, and improves the stability of the optimization process.

Description

Technical Field

[0001] This invention relates to the field of communication chip parameter optimization technology, and in particular to an intelligent optimization method for communication chip design parameters based on deep learning. Background Technology

[0002] As communication systems evolve towards higher speeds, higher integration, and lower power consumption, communication chips need to meet multi-dimensional performance constraints under complex operating conditions. The scale of parameters in the chip design process is constantly expanding, and nonlinear coupling relationships exist between these parameters, significantly increasing design complexity. Traditional communication chip design primarily relies on empirical rules and simulation tools, manually adjusting structural, electrical, and layout parameters iteratively to achieve performance convergence. This approach suffers from low search efficiency, long design cycles, and reliance on design experience. In a multi-parameter, high-dimensional space, the number of parameter combinations grows exponentially. Simply relying on manual adjustment is insufficient to cover the effective design region, easily leading to local optima and impacting overall chip performance.

[0003] In existing methods, some techniques introduce machine learning models to model the mapping relationship between design parameters and performance, thereby reducing the number of simulations. However, existing models are mostly based on static mapping structures, making it difficult to characterize the dynamic correlation and evolution process between parameters, and their adaptability to complex constraints is limited. Furthermore, the parameter optimization process lacks a holistic characterization of the parameter distribution and fails to uniformly model the structural relationships between parameters, resulting in a lack of global consistency in the optimization process. Under multiple constraints, traditional optimization methods typically employ simple boundary restrictions or penalty terms, which cannot finely constrain the parameter variation process, easily leading to parameter oscillations or convergence instability.

[0004] Furthermore, in high-dimensional parameter spaces, existing optimization methods lack effective mechanisms to characterize the importance of parameters, and the differences in the impact of each parameter on performance are not fully utilized, leading to unreasonable resource allocation and limited optimization efficiency. Therefore, there is an urgent need for a communication chip design parameter optimization method that can uniformly describe the parameter relationship structure, characterize the parameter distribution evolution process, and introduce constraint control mechanisms to achieve efficient parameter optimization under complex conditions. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent optimization method for communication chip design parameters based on deep learning. This invention introduces a parameter relationship structure modeling mechanism and a stochastic metric evolution calculation model to perform relationship mapping construction and time series evolution processing on the communication chip design parameters. During the evolution process, entropy flow constraint expression and influence degree driven density modulation mechanism are introduced to perform dynamic control and distributed weighted reconstruction on the parameter update process, forming a continuous optimization parameter iteration process. It has the advantages of controllable parameter search path, high optimization convergence efficiency, and strong parameter distribution consistency under multiple constraints.

[0006] A method for intelligent optimization of communication chip design parameters based on deep learning according to an embodiment of the present invention includes the following steps: Read the communication chip design data and generate a parameter vector sequence and a target performance vector sequence at a unified index location; perform encoding operations on the parameter vector sequence to form a parameter representation vector sequence at the vector dimension location; A deep learning mapping model is constructed at the corresponding positions of the parameter representation vector sequence. Mapping calculations are performed on the parameter representation vector sequence at the node positions of the deep learning mapping model, and a performance prediction vector sequence is generated at the output position. A difference calculation is performed at the corresponding positions of the performance prediction vector sequence and the target performance vector sequence, and an error vector sequence is generated at the difference position. A set of parameter objects is constructed at the corresponding positions of the parameter representation vector sequence. A morphological relationship is established at the object positions. A composite operation is performed at the morphological positions to form a parameter relationship structure. A random measure is generated at the corresponding position of the parameter relationship structure. An evolution calculation is performed on the random measure at the time index position. An entropy flow value is calculated at the evolution position. An entropy flow constraint expression is written at the constraint position. At the parameter update position, the error vector sequence, random measure, and entropy flow constraint value are read and parameter update calculations are performed. An influence degree calculation is performed on the updated parameters at the parameter index position. A density modulation function is generated at the influence degree position. A weighted operation is performed on the random measure at the modulation function position. A modulation parameter vector sequence is generated at the update position. Constraint determination is performed on the modulation parameter vector sequence at the parameter position, and the optimization result of the communication chip design parameters is generated at the output position.

[0007] Optionally, the communication chip design data specifically includes: structural parameter data, electrical parameter data, layout parameter data, process parameter data, operating constraint data, and performance index data.

[0008] Optionally, generating the performance prediction vector sequence at the output position includes the following steps: A deep learning mapping model structure containing input layer node sets, intermediate layer node sets, and output layer node sets is established at the corresponding positions of the parameter representation vector sequence. Connection parameter values ​​are written at the node connection positions. Vector component values ​​of the parameter representation vector sequence are written at the input layer node positions. A numerical mapping is performed between the input layer node index positions and the intermediate layer node index positions, generating an intermediate layer node state sequence at the intermediate layer node positions. A nonlinear transformation operation is performed on the node state values ​​at the intermediate layer node positions, and the transformation result values ​​are written at the transformation positions. Mapping calculations are performed between the intermediate layer node positions and the output layer node positions according to the connection parameter values, generating output node state values ​​at the output layer node positions. Finally, the output node positions are written according to the dimensional order of the target performance vector sequence, generating a performance prediction vector sequence at the output position.

[0009] Optionally, the step of constructing a set of parameter objects at the parameter relationship mapping position, establishing a morphological relationship at the object position, and performing a composite operation at the morphological position to form a parameter relationship structure specifically includes: establishing parameter index numbers at the vector dimension positions of the parameter vector sequence, writing consecutive integer identifiers at the number positions, establishing a mapping relationship between the number positions and the corresponding parameter values, and forming a set of parameter objects at the corresponding positions; constructing a set of parameter index number pairs in the parameter object set according to the parameter index number order, recording two parameter index numbers at the number pair positions, identifying the first parameter index in the number pair as the source parameter index, identifying the second parameter index in the number pair as the target parameter index, writing mapping coefficient values ​​at the number pair positions, and forming a set of morphological relationships at the corresponding positions; multiplying the parameter value at the source parameter index position with the mapping coefficient value at the corresponding position of each parameter index number pair, and writing the product result value at the target parameter index position; accumulating several product result values ​​at the same target parameter index position, and generating a relationship aggregation value at the accumulation position; combining the relationship aggregation value with the original parameter value at each parameter index position, writing the update result value at the combination position, and generating a parameter relationship structure at the corresponding position. Establish a relation record table at each parameter index position in the parameter relation structure, and write the source parameter index, target parameter index and corresponding mapping coefficient value into the record table position to form a parameter relation mapping record structure.

[0010] Optionally, generating a random measure at the corresponding position in the parameter relationship structure and performing evolutionary calculations on the random measure at the time index position includes the following steps: A set of measure positions is established at the parameter index positions of the parameter relationship structure. Each parameter index position is considered a measure position, and parameter relationship values ​​and relationship aggregation values ​​are written to the corresponding positions. The parameter relationship values ​​are mapping coefficient values, forming a measure value sequence at each parameter index position. Normalization calculations are performed on the measure value sequences at each parameter index position, and the normalized result values ​​are written to the corresponding positions, forming a random measure value sequence at each measure position. A time number sequence is established at the time index position, and the random measure value sequence is written to each time number position, forming a measure state sequence at the corresponding time number position. Difference calculations are performed on the measure values ​​at the same parameter index position between adjacent time number positions, and the measure change values ​​are written to the corresponding positions. The measure change values ​​are accumulated at each time number position, and the updated random measure values ​​are written to the corresponding positions, forming a random measure evolution sequence at each time number position. Finally, the random measure evolution sequence is written to the corresponding position of the parameter relationship structure at the time number position, forming an evolved random measure value sequence at the output position.

[0011] Optionally, calculating the entropy flow value at the evolution position and writing the entropy flow constraint expression at the constraint position specifically includes: The random measure evolution sequence is expanded sequentially according to time numbering, forming a sequence structure composed of random measure values ​​corresponding to several time numbers. The random measure value corresponding to each time number is recorded. A logarithmic transformation is performed on the random measure value sequence corresponding to each time number, converting each random measure value into a logarithmic measure value sequence. The random measure value sequence and the logarithmic measure value sequence are matched according to parameter index order, and the values ​​at corresponding positions are multiplied term by term to generate a measure information value sequence. The measure information value sequences under the same time number are summed, and the measure information values ​​corresponding to each parameter index are summed to generate a total measure information value. The difference between the total measure information values ​​corresponding to adjacent time numbers is calculated, and the total measure information value of the current time number is subtracted from the total measure information value of the previous time number to generate an information change value, which is used as the entropy flow value. The entropy flow values ​​are arranged in time number order to form an entropy flow value sequence, which is then recorded. The entropy flow numerical sequence is mapped according to the parameter index order. The entropy flow value corresponding to each time number is written into the corresponding parameter index position. Several entropy flow values ​​corresponding to the same parameter index are accumulated to generate an entropy flow constraint numerical sequence. The entropy flow constraint numerical sequence and the random measure numerical sequence are combined according to the parameter index order. The product and difference of the values ​​at the corresponding positions are jointly calculated to generate a constraint mapping numerical sequence. The entropy flow constraint expression is formed based on the constraint mapping numerical sequence.

[0012] Optionally, the step of performing influence calculation on the updated parameters at the parameter index position and generating the density modulation function at the influence position includes the following steps: The parameter in the parameter index position is the updated parameter value obtained after the parameter update calculation; The updated parameter values ​​generated from the parameter update calculation are expanded according to the parameter index order to form an updated parameter value sequence. This updated parameter value sequence is then matched with the error vector sequence according to the same parameter index order, and the values ​​at corresponding parameter indices are multiplied to form an initial parameter influence value sequence. This initial parameter influence value sequence is normalized to generate a normalized parameter influence value sequence. This normalized parameter influence value sequence is then matched with the random measure value sequence according to the parameter index order, and the values ​​at corresponding parameter indices are weighted to form a weighted parameter influence value sequence. This weighted parameter influence value sequence is then matched with the entropy flow constraint value sequence according to the parameter index order, and the values ​​at corresponding parameter indices are combined to generate a modulated parameter influence value sequence. This modulated parameter influence value sequence is then subjected to exponential mapping calculation to generate a density function value sequence. Finally, this density function value sequence is normalized to generate a density modulation function value sequence.

[0013] Optionally, the constraint determination of the modulation parameter vector sequence at the parameter position specifically includes: The modulation parameter vector sequence is expanded according to the parameter index order to form a modulation parameter value sequence. The modulation parameter value sequence is then matched with the corresponding constraint components in the running constraint data according to the parameter index order to extract the constraint boundary values ​​corresponding to each parameter index. The difference between the modulation parameter values ​​and the corresponding constraint boundary values ​​is calculated to generate a parameter deviation value sequence. The parameter deviation value sequence undergoes sign determination to identify the set of parameter indices that satisfy the constraint range and the set of parameter indices that exceed the constraint range. The modulation parameter values ​​corresponding to parameter indices that satisfy the constraint range are retained and written. The modulation parameter values ​​corresponding to parameter indices that exceed the constraint range undergo boundary correction calculations to adjust the parameter values ​​to the corresponding constraint boundary value range, generating a corrected parameter value sequence. The retained parameter values ​​and the corrected parameter values ​​are combined and written according to the parameter index order to generate a modulation parameter vector sequence that satisfies the constraint conditions.

[0014] The beneficial effects of this invention are: (1) The present invention introduces a parameter object set and morphological relationship construction mechanism at the position of parameter relationship structure, establishes a mapping relationship between parameter indices, forms a relation aggregate value, and forms a structured parameter relationship representation on this basis, so that the correlation between parameters can be uniformly expressed, and the coupling effect between multiple parameters can be considered at the same time during the parameter update process, avoiding the offset problem caused by the independent adjustment of a single parameter, and improving the coordination and overall consistency of parameter combination.

[0015] (2) The present invention constructs a sequence of measure values ​​at the position of random measure evolution, performs evolution calculation in combination with time index, calculates entropy flow values ​​during the evolution process, forms an entropy flow constraint expression, incorporates the parameter change process into constraint control, makes the parameter update process continuous and orderly, reduces the fluctuation range of parameter values, and maintains the stability of parameter distribution structure in multiple iterations, thereby enhancing the controllability of the optimization process.

[0016] (3) The present invention introduces an influence calculation mechanism at the parameter index position, and generates a density modulation function by combining the error vector sequence, the random measure numerical sequence and the entropy flow constraint numerical sequence to perform weighted adjustment on the parameter update process, so that the difference in the role of different parameters in the optimization process can be reflected. Within the same iteration cycle, the parameter components that have a significant impact on performance are adjusted first, thereby improving the parameter update efficiency and accelerating the convergence process under the premise of satisfying the constraint conditions. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for intelligent optimization of communication chip design parameters based on deep learning, as proposed in this invention. Figure 2 This is a schematic diagram of the parameter relationship structure and stochastic metric evolution mechanism of the intelligent optimization method for communication chip design parameters based on deep learning proposed in this invention. Figure 3 This is a schematic diagram of the influence-driven density modulation optimization mechanism of a deep learning-based intelligent optimization method for communication chip design parameters proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-3 A method for intelligent optimization of communication chip design parameters based on deep learning includes the following steps: Read the communication chip design data and generate a parameter vector sequence and a target performance vector sequence at a unified index location; perform encoding operations on the parameter vector sequence to form a parameter representation vector sequence at the vector dimension location; A deep learning mapping model is constructed at the corresponding positions of the parameter representation vector sequence. Mapping calculations are performed on the parameter representation vector sequence at the node positions of the deep learning mapping model, and a performance prediction vector sequence is generated at the output position. A difference calculation is performed at the corresponding positions of the performance prediction vector sequence and the target performance vector sequence, and an error vector sequence is generated at the difference position. A set of parameter objects is constructed at the corresponding positions of the parameter representation vector sequence. A morphological relationship is established at the object positions. A composite operation is performed at the morphological positions to form a parameter relationship structure. A random measure is generated at the corresponding position of the parameter relationship structure. An evolution calculation is performed on the random measure at the time index position. An entropy flow value is calculated at the evolution position. An entropy flow constraint expression is written at the constraint position. At the parameter update position, the error vector sequence, random measure, and entropy flow constraint value are read and parameter update calculations are performed. An influence degree calculation is performed on the updated parameters at the parameter index position. A density modulation function is generated at the influence degree position. A weighted operation is performed on the random measure at the modulation function position. A modulation parameter vector sequence is generated at the update position. Constraint determination is performed on the modulation parameter vector sequence at the parameter position, and the optimization result of the communication chip design parameters is generated at the output position.

[0020] In this embodiment, the communication chip design data specifically includes: structural parameter data, electrical parameter data, layout parameter data, process parameter data, operating constraint data, and performance index data.

[0021] In this embodiment, generating a performance prediction vector sequence at the output position includes the following steps: A deep learning mapping model structure containing input layer node sets, intermediate layer node sets, and output layer node sets is established at the corresponding positions of the parameter representation vector sequence. Connection parameter values ​​are written at the node connection positions. Vector component values ​​of the parameter representation vector sequence are written at the input layer node positions. A numerical mapping is performed between the input layer node index positions and the intermediate layer node index positions, generating an intermediate layer node state sequence at the intermediate layer node positions. A nonlinear transformation operation is performed on the node state values ​​at the intermediate layer node positions, and the transformation result values ​​are written at the transformation positions. Mapping calculations are performed between the intermediate layer node positions and the output layer node positions according to the connection parameter values, generating output node state values ​​at the output layer node positions. Finally, the output node positions are written according to the dimensional order of the target performance vector sequence, generating a performance prediction vector sequence at the output position.

[0022] In this embodiment, constructing a parameter object set at the parameter relationship mapping position, establishing a morphic relationship at the object position, and performing a composite operation at the morphic position to form a parameter relationship structure specifically includes: establishing parameter index numbers at the vector dimension positions of the parameter vector sequence, writing consecutive integer identifiers at the number positions, establishing a mapping relationship between the number positions and the corresponding parameter values, and forming a parameter object set at the corresponding positions; constructing a parameter index number pair set in the parameter object set according to the parameter index number order, recording two parameter index numbers at the number pair positions, identifying the first parameter index in the number pair as the source parameter index, identifying the second parameter index in the number pair as the target parameter index, writing the mapping coefficient value at the number pair positions, and forming a morphic relationship set at the corresponding positions; multiplying the parameter value at the source parameter index position with the mapping coefficient value at the corresponding position of each parameter index number pair, and writing the product result value at the target parameter index position; accumulating several product result values ​​at the same target parameter index position, and generating a relationship aggregation value at the accumulation position; combining the relationship aggregation value with the original parameter value at each parameter index position, writing the update result value at the combination position, and generating a parameter relationship structure at the corresponding position. Establish a relation record table at each parameter index position in the parameter relation structure, and write the source parameter index, target parameter index and corresponding mapping coefficient value into the record table position to form a parameter relation mapping record structure.

[0023] In this embodiment, generating a random measure at the corresponding position in the parameter relationship structure and performing evolutionary calculations on the random measure at the time index position includes the following steps: A set of measure positions is established at the parameter index positions of the parameter relationship structure. Each parameter index position is considered a measure position, and parameter relationship values ​​and relationship aggregation values ​​are written to the corresponding positions. The parameter relationship values ​​are mapping coefficient values, forming a measure value sequence at each parameter index position. Normalization calculations are performed on the measure value sequences at each parameter index position, and the normalized result values ​​are written to the corresponding positions, forming a random measure value sequence at each measure position. A time number sequence is established at the time index position, and the random measure value sequence is written to each time number position, forming a measure state sequence at the corresponding time number position. Difference calculations are performed on the measure values ​​at the same parameter index position between adjacent time number positions, and the measure change values ​​are written to the corresponding positions. The measure change values ​​are accumulated at each time number position, and the updated random measure values ​​are written to the corresponding positions, forming a random measure evolution sequence at each time number position. Finally, the random measure evolution sequence is written to the corresponding position of the parameter relationship structure at the time number position, forming an evolved random measure value sequence at the output position.

[0024] In this embodiment, calculating the entropy flow value at the evolution position and writing the entropy flow constraint expression at the constraint position specifically includes: The random measure evolution sequence is expanded sequentially according to time numbering, forming a sequence structure composed of random measure values ​​corresponding to several time numbers. The random measure value corresponding to each time number is recorded. A logarithmic transformation is performed on the random measure value sequence corresponding to each time number, converting each random measure value into a logarithmic measure value sequence. The random measure value sequence and the logarithmic measure value sequence are matched according to parameter index order, and the values ​​at corresponding positions are multiplied term by term to generate a measure information value sequence. The measure information value sequences under the same time number are summed, and the measure information values ​​corresponding to each parameter index are summed to generate a total measure information value. The difference between the total measure information values ​​corresponding to adjacent time numbers is calculated, and the total measure information value of the current time number is subtracted from the total measure information value of the previous time number to generate an information change value, which is used as the entropy flow value. The entropy flow values ​​are arranged in time number order to form an entropy flow value sequence, which is then recorded. The entropy flow numerical sequence is mapped according to the parameter index order. The entropy flow value corresponding to each time number is written into the corresponding parameter index position. Several entropy flow values ​​corresponding to the same parameter index are accumulated to generate an entropy flow constraint numerical sequence. The entropy flow constraint numerical sequence and the random measure numerical sequence are combined according to the parameter index order. The product and difference of the values ​​at the corresponding positions are jointly calculated to generate a constraint mapping numerical sequence. The entropy flow constraint expression is formed based on the constraint mapping numerical sequence.

[0025] In this embodiment, performing influence calculation on the updated parameters at the parameter index position and generating the density modulation function at the influence position includes the following steps: The parameter in the parameter index position is the updated parameter value obtained after the parameter update calculation; The updated parameter values ​​generated from the parameter update calculation are expanded according to the parameter index order to form an updated parameter value sequence. This updated parameter value sequence is then matched with the error vector sequence according to the same parameter index order, and the values ​​at corresponding parameter indices are multiplied to form an initial parameter influence value sequence. This initial parameter influence value sequence is normalized to generate a normalized parameter influence value sequence. This normalized parameter influence value sequence is then matched with the random measure value sequence according to the parameter index order, and the values ​​at corresponding parameter indices are weighted to form a weighted parameter influence value sequence. This weighted parameter influence value sequence is then matched with the entropy flow constraint value sequence according to the parameter index order, and the values ​​at corresponding parameter indices are combined to generate a modulated parameter influence value sequence. This modulated parameter influence value sequence is then subjected to exponential mapping calculation to generate a density function value sequence. Finally, this density function value sequence is normalized to generate a density modulation function value sequence.

[0026] In this embodiment, the constraint determination of the modulation parameter vector sequence at the parameter position specifically includes: The modulation parameter vector sequence is expanded according to the parameter index order to form a modulation parameter value sequence. The modulation parameter value sequence is then matched with the corresponding constraint components in the running constraint data according to the parameter index order to extract the constraint boundary values ​​corresponding to each parameter index. The difference between the modulation parameter values ​​and the corresponding constraint boundary values ​​is calculated to generate a parameter deviation value sequence. The parameter deviation value sequence undergoes sign determination to identify the set of parameter indices that satisfy the constraint range and the set of parameter indices that exceed the constraint range. The modulation parameter values ​​corresponding to parameter indices that satisfy the constraint range are retained and written. The modulation parameter values ​​corresponding to parameter indices that exceed the constraint range undergo boundary correction calculations to adjust the parameter values ​​to the corresponding constraint boundary value range, generating a corrected parameter value sequence. The retained parameter values ​​and the corrected parameter values ​​are combined and written according to the parameter index order to generate a modulation parameter vector sequence that satisfies the constraint conditions.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to a communication chip design environment. This environment contains a large number of structural parameters, electrical parameters, layout parameters, and process parameters, with complex coupling relationships between different parameters. During the design process, it is often necessary to balance various performance indicators, such as power consumption, latency, signal integrity, and stability. Traditional methods mainly rely on manual experience for parameter adjustment, combined with continuous iteration using simulation tools. This often requires a long period to obtain a satisfactory parameter combination and is prone to parameter oscillation and local convergence problems. In this environment, the communication chip design data is first read, and a parameter vector sequence and a target performance vector sequence are generated at a unified index location. Subsequently, encoding operations are performed on the parameter vector sequence to form a parameter representation vector sequence at the vector dimension location. Based on this parameter representation vector sequence, a deep learning mapping model is constructed at the corresponding location, and mapping calculations are performed at the model node locations to generate a performance prediction vector sequence. By calculating the corresponding difference between the performance prediction vector sequence and the target performance vector sequence, an error vector sequence is generated, providing a quantitative basis for subsequent parameter adjustments. Furthermore, a set of parameter objects is constructed at the corresponding positions of the parameter representation vector sequence, morphological relationships are established between the objects, and a parameter relationship structure is formed through composition operations. Based on this structure, a random measure is generated, and evolutionary calculations are performed on the random measure at the time index position. By cumulatively analyzing the measure changes during the evolution process, entropy flow values ​​are calculated, and entropy flow constraint expressions are written at the constraint positions to ensure the continuity and directionality of the parameter change process. During the parameter update phase, the error vector sequence, random measure, and entropy flow constraint values ​​are read, and update calculations are performed on the parameters. Subsequently, the influence degree is calculated on the updated parameters at the parameter index position, and a density modulation function is generated by combining it with the error vector sequence and measure information. This function performs weighted operations on the random measure to generate a modulation parameter vector sequence. Finally, constraint judgment is performed on the modulation parameter vector sequence at the parameter position, and parameters that do not meet the constraint range are corrected at the boundary, outputting the optimized parameters of the communication chip design that meet the constraint conditions.

[0028] Table 1: Comparison of Optimization Effects of Communication Chip Design Parameters

[0029] As shown in Table 1, in terms of power consumption, the traditional method yielded 125mW, while the result using this invention was reduced to 98mW, a decrease of 27mW, corresponding to a reduction of approximately 21.6%. Regarding latency, it decreased from 2.8ns to 2.1ns, a reduction of 0.7ns, representing a decrease of 25.0%. In terms of signal integrity error, it decreased from 0.18 to 0.11, a reduction of 0.07, corresponding to a reduction of approximately 38.9%. Stability fluctuation decreased from 0.26 to 0.09, a decrease of 0.17, representing an improvement of 65.4%, indicating a significant reduction in fluctuation during parameter evolution. Regarding parameter convergence error, it decreased from 0.12 to 0.05, a reduction of 0.07, representing a reduction of 58.3%, indicating a more accurate parameter update process. In terms of iteration count, the traditional method typically requires more than 300 iterations, while this method controls the number of iterations to around 290 across multiple metrics, reducing the number by approximately 10 to 30 iterations. In terms of convergence time, it was reduced from 48 minutes to 42 minutes, a reduction of approximately 6 minutes, or about 12.5%. Regarding the overall design cycle, it was reduced from 72 hours to 38 hours, a reduction of 34 hours, or about 47.2%. These data demonstrate that this method exhibits significant advantages across multiple dimensions, achieving superior performance metrics and demonstrating higher levels of computational and convergence efficiency. Particularly in terms of stability fluctuation amplitude and parameter convergence error, the improvement rate exceeds 50%, indicating that the introduction of parameter relationship structures, stochastic metric evolution, and entropy flow constraints makes the parameter update process smoother and more directional. Furthermore, the density modulation function differentiates the degree of influence of different parameters, allowing key parameters to be adjusted preferentially, thereby reaching the target performance range within a fewer iterations. Overall, this method demonstrates good adaptability and stability in complex multi-parameter optimization scenarios.

[0030] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent optimization of communication chip design parameters based on deep learning, characterized in that, Includes the following steps: Read the communication chip design data and generate a parameter vector sequence and a target performance vector sequence at a unified index location; perform encoding operations on the parameter vector sequence to form a parameter representation vector sequence at the vector dimension location; A deep learning mapping model is constructed at the corresponding positions of the parameter representation vector sequence. Mapping calculations are performed on the parameter representation vector sequence at the node positions of the deep learning mapping model, and a performance prediction vector sequence is generated at the output position. A difference calculation is performed at the corresponding positions of the performance prediction vector sequence and the target performance vector sequence, and an error vector sequence is generated at the difference position. A set of parameter objects is constructed at the corresponding positions of the parameter representation vector sequence. A morphological relationship is established at the object positions. A composite operation is performed at the morphological positions to form a parameter relationship structure. A random measure is generated at the corresponding position of the parameter relationship structure. An evolution calculation is performed on the random measure at the time index position. An entropy flow value is calculated at the evolution position. An entropy flow constraint expression is written at the constraint position. At the parameter update position, the error vector sequence, random measure, and entropy flow constraint value are read and parameter update calculations are performed. An influence degree calculation is performed on the updated parameters at the parameter index position. A density modulation function is generated at the influence degree position. A weighted operation is performed on the random measure at the modulation function position. A modulation parameter vector sequence is generated at the update position. Constraint determination is performed on the modulation parameter vector sequence at the parameter position, and the optimization result of the communication chip design parameters is generated at the output position.

2. The intelligent optimization method for communication chip design parameters based on deep learning according to claim 1, characterized in that, The communication chip design data specifically includes: structural parameter data, electrical parameter data, layout parameter data, process parameter data, operating constraint data, and performance index data.

3. The intelligent optimization method for communication chip design parameters based on deep learning according to claim 2, characterized in that, The process of generating a performance prediction vector sequence at the output position includes the following steps: A deep learning mapping model structure containing input layer node sets, intermediate layer node sets, and output layer node sets is established at the corresponding positions of the parameter representation vector sequence. Connection parameter values ​​are written at the node connection positions. Vector component values ​​of the parameter representation vector sequence are written at the input layer node positions. A numerical mapping is performed between the input layer node index positions and the intermediate layer node index positions, generating an intermediate layer node state sequence at the intermediate layer node positions. A nonlinear transformation operation is performed on the node state values ​​at the intermediate layer node positions, and the transformation result values ​​are written at the transformation positions. Mapping calculations are performed between the intermediate layer node positions and the output layer node positions according to the connection parameter values, generating output node state values ​​at the output layer node positions. Finally, the output node positions are written according to the dimensional order of the target performance vector sequence, generating a performance prediction vector sequence at the output position.

4. The intelligent optimization method for communication chip design parameters based on deep learning according to claim 3, characterized in that, The specific steps of constructing a parameter object set at the parameter relationship mapping position, establishing a morphological relationship at the object position, and performing a composite operation at the morphological position to form a parameter relationship structure include: Establishing parameter index numbers at the vector dimension positions of the parameter vector sequence, writing consecutive integer identifiers at the index positions, establishing a mapping relationship between the index positions and corresponding parameter values, and forming a parameter object set at the corresponding positions; Constructing a set of parameter index number pairs in the parameter object set according to the parameter index number order, recording two parameter index numbers at the index pair positions, identifying the first parameter index in the index pair as the source parameter index, identifying the second parameter index in the index pair as the target parameter index, writing mapping coefficient values ​​at the index pair positions, and forming a morphological relationship set at the corresponding positions; Multiplying the parameter value at the source parameter index position with the mapping coefficient value at the corresponding position of each parameter index number pair, and writing the product result value at the target parameter index position; Accumulating several product result values ​​at the same target parameter index position, and generating a relationship aggregation value at the accumulation position; Combining the relationship aggregation value with the original parameter value at each parameter index position, writing the update result value at the combination position, and generating a parameter relationship structure at the corresponding position. Establish a relation record table at each parameter index position in the parameter relation structure, and write the source parameter index, target parameter index and corresponding mapping coefficient value into the record table position to form a parameter relation mapping record structure.

5. The intelligent optimization method for communication chip design parameters based on deep learning according to claim 4, characterized in that, The process of generating a random measure at the corresponding position in the parameter relationship structure and performing evolutionary calculations on the random measure at the time index position includes the following steps: A set of measure positions is established at the parameter index positions of the parameter relationship structure. Each parameter index position is designated as a measure position, and parameter relationship values ​​and relationship aggregation values ​​are written to the corresponding positions. The parameter relationship values ​​are mapping coefficient values, forming a measure value sequence at each parameter index position. Normalization calculations are performed on the measure value sequences at each parameter index position, and the normalized result values ​​are written to the corresponding positions, forming a random measure value sequence at each measure position. A time number sequence is established at the time index position, and the random measure value sequence is written to each time number position, forming a measure state sequence at the corresponding time number position. Difference calculations are performed on the measure values ​​at the same parameter index position between adjacent time number positions, and the measure change values ​​are written to the corresponding positions. The measure change values ​​are accumulated at each time number position, and the updated random measure values ​​are written to the corresponding positions, forming a random measure evolution sequence at each time number position. Finally, the random measure evolution sequence is written to the corresponding position in the parameter relationship structure at the final time number position, forming an evolved random measure value sequence at the output position.

6. The intelligent optimization method for communication chip design parameters based on deep learning according to claim 5, characterized in that, The calculation of entropy flow values ​​at evolutionary positions and the writing of entropy flow constraint expressions at constraint positions specifically include: The random measure evolution sequence is expanded sequentially according to time numbering, forming a sequence structure composed of random measure values ​​corresponding to several time numbers. The random measure value corresponding to each time number is recorded. A logarithmic transformation is performed on the random measure value sequence corresponding to each time number, converting each random measure value into a logarithmic measure value sequence. The random measure value sequence and the logarithmic measure value sequence are matched according to parameter index order, and the values ​​at corresponding positions are multiplied term by term to generate a measure information value sequence. The measure information value sequences under the same time number are summed, and the measure information values ​​corresponding to each parameter index are summed to generate a total measure information value. The difference between the total measure information values ​​corresponding to adjacent time numbers is calculated, and the total measure information value of the current time number is subtracted from the total measure information value of the previous time number to generate an information change value, which is used as the entropy flow value. The entropy flow values ​​are arranged in time number order to form an entropy flow value sequence, which is then recorded. The entropy flow numerical sequence is mapped according to the parameter index order. The entropy flow value corresponding to each time number is written into the corresponding parameter index position. Several entropy flow values ​​corresponding to the same parameter index are accumulated to generate an entropy flow constraint numerical sequence. The entropy flow constraint numerical sequence and the random measure numerical sequence are combined according to the parameter index order. The product and difference of the values ​​at the corresponding positions are jointly calculated to generate a constraint mapping numerical sequence. The entropy flow constraint expression is formed based on the constraint mapping numerical sequence.

7. The intelligent optimization method for communication chip design parameters based on deep learning according to claim 6, characterized in that, The step of performing influence calculation on the updated parameters at the parameter index position and generating the density modulation function at the influence position includes the following steps: The parameter in the parameter index position is the updated parameter value obtained after the parameter update calculation; The updated parameter values ​​generated from the parameter update calculation are expanded according to the parameter index order to form an updated parameter value sequence. This updated parameter value sequence is then matched with the error vector sequence according to the same parameter index order, and the values ​​at corresponding parameter indices are multiplied to form an initial parameter influence value sequence. This initial parameter influence value sequence is normalized to generate a normalized parameter influence value sequence. This normalized parameter influence value sequence is then matched with the random measure value sequence according to the parameter index order, and the values ​​at corresponding parameter indices are weighted to form a weighted parameter influence value sequence. This weighted parameter influence value sequence is then matched with the entropy flow constraint value sequence according to the parameter index order, and the values ​​at corresponding parameter indices are combined to generate a modulated parameter influence value sequence. This modulated parameter influence value sequence is then subjected to exponential mapping calculation to generate a density function value sequence. Finally, this density function value sequence is normalized to generate a density modulation function value sequence.

8. The intelligent optimization method for communication chip design parameters based on deep learning according to claim 7, characterized in that, The constraint determination of the modulation parameter vector sequence at the parameter position specifically includes: The modulation parameter vector sequence is expanded according to the parameter index order to form a modulation parameter value sequence. The modulation parameter value sequence is then matched with the corresponding constraint components in the running constraint data according to the parameter index order to extract the constraint boundary values ​​corresponding to each parameter index. The difference between the modulation parameter values ​​and the corresponding constraint boundary values ​​is calculated to generate a parameter deviation value sequence. The parameter deviation value sequence undergoes sign determination to identify the set of parameter indices that satisfy the constraint range and the set of parameter indices that exceed the constraint range. The modulation parameter values ​​corresponding to the parameter indices that satisfy the constraint range are retained and written. The modulation parameter values ​​corresponding to the parameter indices that exceed the constraint range undergo boundary correction calculations to adjust the parameter values ​​to the corresponding constraint boundary value range, generating a corrected parameter value sequence. The retained parameter values ​​and the corrected parameter values ​​are combined and written according to the parameter index order to generate a modulation parameter vector sequence that satisfies the constraint conditions.

Images