Cost controlled phased resolution generation and cache reuse system and method

Abstract

The application discloses a cost-controlled phased resolution generation and cache reuse system and method applied to the field of computer technology, which firstly performs rapid preliminary screening at a low resolution level, and removes candidate configurations which are obviously unreasonable in physical structure or whose energy significantly deviates from a threshold; then gradually improves the resolution of the configurations, introduces a physical constraint and an energy calculation model, and gradually approaches a high-fidelity simulation result; meanwhile, a structure feature cache mechanism is arranged at each resolution level, intermediate states and energy gradient information of the calculated configurations are standardized stored and indexed, and are called at a higher resolution level for subsequent same or similar configurations, so that repeated calculation is avoided; finally, through a resolution transition determination module, whether to enter a next resolution level or terminate calculation is dynamically decided according to an energy convergence trend of the current configuration, a structure stability index and a calculation resource budget, so that an optimal balance between physical rationality and calculation efficiency is realized in a global range.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a cost-controlled staged resolution generation and cache reuse system and method. Background Technology

[0002] With the deep penetration of computer graphics and physics simulation technologies into fields such as film and television special effects, game engines, and industrial digital twins, the demand for real-time generation of high-fidelity dynamic scenes continues to rise. Traditional physics simulation methods rely on full-resolution meshes and continuous time-step solutions, and their core advantage lies in their ability to accurately reproduce complex physical behaviors such as material deformation, fluid motion, or rigid body collisions. However, such methods face severe computational bottlenecks in large-scale scenes or multi-object interactions: to maintain physical plausibility, the system needs to iteratively calculate global geometric details and dynamic constraints in each frame, resulting in an exponential increase in single-frame latency, and a sharp increase in memory usage with increasing resolution, severely restricting their deployment feasibility in real-time interactive or high-throughput screening scenarios.

[0003] Among these, phased resolution generation and intermediate state caching reuse techniques are considered key paths to alleviate computational pressure. The basic idea behind this technique is to construct a progressive computational process from low-precision coarse simulation to high-precision detail supplementation, thereby eliminating redundant computational load while ensuring final visual and physical consistency. However, existing implementations generally suffer from fragmented strategies and inefficient resource scheduling: the coarse-grained stage lacks constraints and guidance on key physical energy nodes, leading to repeated corrections of trajectory deviations in subsequent refinement stages; intermediate computational results lack structured caching mechanisms and batch processing scheduling, making cross-frame reuse impossible in similar scenarios or parameter fine-tuning; and memory management strategies are not dynamically coupled with computational complexity, often triggering resource overflows or forced degradation due to peak usage, thus compromising physical consistency.

[0004] Existing technologies face a structural contradiction between pursuing physical fidelity and controlling computational costs: on the one hand, while full-resolution continuous simulation can ensure physical plausibility, its high computational density and memory requirements make it difficult to support large-scale parallel screening or real-time interaction; on the other hand, existing staged or caching strategies suffer from uncontrollable accuracy loss or unstable acceleration gains due to a lack of energy constraint guidance, coarse reuse granularity, and rigid resource scheduling. Especially in scenarios such as industrial design and special effects pre-visualization that require high-frequency parameter tuning and comparison of multiple schemes, existing methods struggle to achieve a predictable balance between throughput efficiency, memory usage, and physical accuracy. There is an urgent need for a staged resolution generation and caching co-optimization scheme that can achieve controlled computational costs, constrained accuracy, and reusable resources. Summary of the Invention

[0005] The core of this invention lies in constructing a multi-level resolution collaborative evolution architecture. At each level, strict energy screening thresholds and structural feature caching and reuse mechanisms are set. By controlling the investment of computing resources in stages, efficient computing in large-scale screening scenarios is achieved while ensuring physical rationality. This solves the technical problems of excessively high cost and serious resource waste in high-precision computing in large-scale molecular configuration screening in existing technologies. At the same time, the caching and reuse mechanism runs through the entire process, allowing similar configurations to directly call historical results or perform interpolation prediction in high-level calculations, which significantly reduces the overhead of repeated calculations.

[0006] To solve the above problems, the present invention adopts the following technical solution.

[0007] A cost-controlled, phased resolution generation and cache reuse method includes the following steps: Step S1: Receive the initial molecular configuration set to be evaluated through the configuration input interface module. The initial molecular configuration set contains multiple three-dimensional spatial coordinate sequences to be screened. Step S2: At the first resolution level, for each configuration in the initial molecular configuration set, the simplified potential energy function is used to calculate its coarse-grained energy value. The simplified potential energy function includes only bond length, bond angle and van der Waals repulsion term. Step S3: Based on the coarse-grained energy value, set a first energy threshold, remove configurations with energy values ​​higher than the first energy threshold, and retain the remaining configurations to form a first set of screened configurations; Step S4: For each configuration in the first selection subset, extract its atomic arrangement topological features and local geometric invariants, generate a first-level structural feature vector, and associate the first-level structural feature vector with the corresponding coarse-grained energy value and store it in the first-level cache unit. Step S5: At the second resolution level, for each configuration in the first selected configuration subset, call the first-level structure feature vector stored in the first-level cache unit to initialize the fine potential energy function calculation process. The fine potential energy function includes all-atom force field parameters, long-range Coulomb interaction terms, and implicit solvent model. Step S6: Perform fine potential energy function calculation to obtain the second-level high-precision energy value and energy gradient vector; Step S7: Based on the second-level high-precision energy value, set a second energy threshold, remove configurations with energy values ​​higher than the second energy threshold, and retain the remaining configurations to form a second set of filtered configurations; Step S8: For each configuration in the second set of selected configurations, extract its full atomic coordinate offset, bond angle distribution histogram and solvent accessible surface area change rate to generate a second-level structure feature vector, and store the second-level structure feature vector, high-precision energy value and energy gradient vector together in the second-level cache unit. Step S9: At the third resolution level, for each configuration in the second selection configuration subset, call the second-level structure feature vector and energy gradient vector stored in the second-level cache unit as initial conditions input to the quantum mechanical semi-empirical method calculation module. The quantum mechanical semi-empirical method uses parameterized Hamiltonian to approximate the electronic structure. Step S10: Perform quantum mechanical semi-empirical method calculations to obtain the third-level electron energy correction value and orbital occupancy distribution; Step S11: Superimpose the third-level electron energy correction value onto the second-level high-precision energy value to generate the final comprehensive energy assessment result; Step S12: Based on the final comprehensive energy assessment results, set the final energy ranking threshold, and output the configurations with energy values ​​lower than the final energy ranking threshold as the preferred configuration set that meets the physical rationality standard and has controllable computational cost.

[0008] Furthermore, the specific operations of step S2 include the following: mapping the three-dimensional spatial coordinate sequence to a preset coarse-grained grid system, with the grid spacing ranging from 0.5 nanometers to 1.5 nanometers; performing single-point energy calculations on the mapped configuration, simplifying the potential energy function by using the harmonic oscillator potential model for the bond length term, the cosine expansion potential model for the bond angle term, and the inverse 12th power potential model for the van der Waals repulsion term; all model parameters are extracted and fixed from the general biomolecular force field parameter set, without dynamic fitting.

[0009] Furthermore, the generation operation of the first-level structural feature vector in step S4 includes the following: calculating the local density, average coordination number and spatial distribution entropy of each coarse-grained unit within the configuration, and combining the local density, average coordination number and spatial distribution entropy into the first-level structural feature vector, which is used for subsequent similarity matching and cache retrieval.

[0010] Furthermore, the first-level cache unit adopts a hash table data structure, using a normalized floating-point array of the first-level structure feature vectors as the key and a coarse-grained energy value as the storage value, supporting feature matching and energy value retrieval with constant time complexity.

[0011] Further, the specific operations of step S6 include the following: restoring the original atomic-level coordinates of the configuration, loading the full atomic force field parameter library, which covers the AMBER, CHARMM and OPLS series force fields; using the conjugate gradient algorithm to iteratively optimize the atomic positions until the energy gradient norm converges to below 0.001 electron volts per angstrom; and simultaneously recording the atomic coordinates, energy values ​​and gradient vectors during the optimization process.

[0012] Furthermore, the generation operation of the second-level structure feature vector in step S8 includes the following: calculating the Euclidean norm of the atomic coordinate offsets before and after optimization; statistically analyzing the bond angle distribution histogram, dividing it into intervals of ten degrees; calculating the absolute value of the rate of change of solvent-accessible surface area; and combining the Euclidean norm, the bond angle distribution histogram, and the absolute value of the rate of change of solvent-accessible surface area into the second-level structure feature vector.

[0013] Furthermore, the second-level cache unit uses a key-value pair database. The key consists of a low-dimensional code generated by principal component analysis to reduce the dimensionality of the second-level structural feature vector. The value contains a high-precision energy value, an energy gradient vector, and a pointer to the original configuration identifier, supporting energy prediction of nearest neighbor configurations based on similarity metrics.

[0014] Furthermore, the specific operation of step S10 includes the following: based on the gradient sensitivity analysis results in the second-level buffer, quantum correction is performed only on atoms with gradient magnitudes greater than 0.5 electron volts per angstrom and their neighboring atoms within the third order; the classical force field energy is used for the remaining regions; the total energy after correction is composed of the superposition of the quantum region energy, the non-quantum region energy and the buffer layer coupling energy.

[0015] Furthermore, the buffer layer coupling energy can be smoothly transitioned through a linear interpolation function, ensuring that the energy function is continuously differentiable throughout the entire domain. The interpolation weight of the linear interpolation function is determined by the distance of the atom from the boundary of the quantum region.

[0016] Furthermore, the quantum mechanical semi-empirical method calculation module adopts a block diagonalization strategy, dividing the molecular system into multiple subsystems according to chemical bond connectivity. The electronic Hamiltonian matrix is ​​solved independently for each subsystem, and the coupling between subsystems is corrected by the boundary atomic orbital overlap integral, which reduces the matrix solution dimension and improves the computational efficiency.

[0017] Furthermore, the configuration input interface module supports parsing standard molecular file formats, including but not limited to three-dimensional coordinate files, connectivity description files, and charge distribution files. During import, it automatically performs atom type mapping, missing hydrogen atom completion, and initial geometry optimization to ensure that the input configuration meets the format and physical consistency requirements of subsequent calculation modules.

[0018] A cost-controlled phased resolution generation and cache reuse system, applied to the aforementioned cost-controlled phased resolution generation and cache reuse method, includes the following modules: Configuration input interface module, used to receive the initial set of molecular configurations to be evaluated; The first resolution-level calculation module is used to calculate the coarse-grained energy value of each configuration in the initial molecular configuration set using a simplified potential energy function; The first-level filtering module is used to filter configurations that meet the requirements and form a first set of filtered configurations. The first-level feature extraction and caching module is used to generate the first-level structural feature vector and associate the first-level structural feature vector with the corresponding coarse-grained energy value and store it in the first-level cache unit. The second resolution level calculation module is used to obtain the high-precision energy value and energy gradient vector of the second level; The second-level filtering module is used to filter configurations that meet the requirements and form a second set of filtered configurations. The second-level feature extraction and caching module is used to generate the second-level structural feature vector and store the second-level structural feature vector, high-precision energy value and energy gradient vector together in the second-level cache unit. The third-resolution level calculation module is used to obtain the third-level electron energy correction value and orbital occupancy distribution, and generate the final comprehensive energy assessment result. The final configuration output module is used to output configurations with energy values ​​lower than the final energy ranking threshold as a set of preferred configurations that meet physical rationality standards and have controllable computational costs.

[0019] Furthermore, it also includes a resolution transition determination module, which is integrated between the second resolution level calculation module and the third resolution level calculation module. Its inputs are the energy gradient magnitude, bond angle standard deviation, and absolute value of the solvent-accessible surface area change rate calculated at the second level for the current configuration. Its output is a Boolean decision signal. When the energy gradient magnitude is less than the preset convergence threshold, the bond angle standard deviation is less than the structural stability threshold, and the absolute value of the solvent-accessible surface area change rate is less than the environmental disturbance tolerance threshold, it is determined that the configuration does not need to enter the third resolution level, and the high-precision energy value of the second level is directly used as the final evaluation result.

[0020] Compared with the prior art, the advantages of this invention are: (1) This invention constructs a phased resolution evolution architecture, decomposing the computationally expensive physical rationality assessment process into multiple controllable levels. Each level sets strict energy screening thresholds and structural feature caching mechanisms, effectively avoiding ineffective high-precision calculations for low-potential configurations. The first level uses a simplified potential energy function to achieve millisecond-level initial screening, eliminating more than 80% of obviously unreasonable configurations. The second level introduces an all-atom force field and solvent model to further screen the remaining configurations, while caching structural features and energy gradients to provide high-quality initial conditions for the third level. The third level only performs quantum mechanical semi-empirical calculations on the most promising configurations, and reduces computational complexity through a block diagonalization strategy.

[0021] (2) The cache reuse mechanism runs through the entire process, allowing similar configurations to directly call historical results or perform interpolation predictions in high-level calculations, significantly reducing the overhead of repeated calculations. The resolution transition determination module dynamically terminates the calculation based on energy convergence and structural stability, avoiding excessive resource investment. The overall scheme reduces the average computation cost to less than five percent of the traditional full quantum computing method while ensuring that the physical rationality of the final output configuration is no less than that of existing high-precision methods, while increasing the screening throughput by two orders of magnitude. It is suitable for large-scale computing scenarios such as virtual screening of drug lead compounds, prediction of material crystal structures, and protein conformation sampling. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a logical flowchart of the first resolution level in this invention: coarse-grained initial screening and structural feature caching; Figure 3 This is a flowchart illustrating the logical flow of the second resolution level in this invention: full-atom force field precision calculation and gradient cache reuse. Figure 4 This is a logical flowchart of the third resolution level in this invention: quantum semi-empirical correction and energy superposition decision-making; Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow between the cross-level caching system and the similar configuration feature matching mechanism in this invention. Detailed Implementation

[0023] The technical solutions will now be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention.

[0024] Example:

[0025] Please see Figures 1-5 A cost-controlled, phased resolution generation and cache reuse method includes the following steps: Step S1: Receive the initial molecular configuration set to be evaluated through the configuration input interface module. The initial molecular configuration set contains multiple three-dimensional spatial coordinate sequences to be screened. The configuration input interface module supports parsing standard molecular file formats, including but not limited to three-dimensional coordinate files, connectivity description files, and charge distribution files. During import, it automatically performs atom type mapping, missing hydrogen atom completion, and initial geometry optimization to ensure that the input configuration meets the format and physical consistency requirements of subsequent calculation modules.

[0026] Step S2: At the first resolution level, for each configuration in the initial molecular configuration set, the simplified potential energy function is used to calculate its coarse-grained energy value. The simplified potential energy function includes only bond length, bond angle and van der Waals repulsion terms, ignoring long-range electrostatic interactions and solvation effects. Specifically: the three-dimensional spatial coordinate sequence is mapped to a preset coarse-grained grid system, with the grid spacing ranging from 0.5 nanometers to 1.5 nanometers; single-point energy calculations are performed on the mapped configurations, and the bond length term in the simplified potential function adopts the harmonic oscillator potential model, the bond angle term adopts the cosine expansion potential model, and the van der Waals repulsion term adopts the twelfth power inverse proportional potential model; all model parameters are extracted and fixed from the general biomolecular force field parameter set, and no dynamic fitting is performed.

[0027] Step S3: Based on the coarse-grained energy value, set a first energy threshold, remove configurations with energy values ​​higher than the first energy threshold, and retain the remaining configurations to form a first set of screened configurations; Step S4: For each configuration in the first selection subset, extract its atomic arrangement topological features and local geometric invariants, generate a first-level structural feature vector, and associate the first-level structural feature vector with the corresponding coarse-grained energy value and store it in the first-level cache unit. Specifically, the generation of the first-level structural feature vector includes the following steps: calculating the local density, average coordination number, and spatial distribution entropy of each coarse-grained unit within the configuration; combining the local density, average coordination number, and spatial distribution entropy into the first-level structural feature vector; and using the first-level structural feature vector for subsequent similarity matching and cache retrieval. The first-level cache unit adopts a hash table data structure, using a normalized floating-point array of the first-level structure feature vectors as keys and coarse-grained energy values ​​as storage values, supporting feature matching and energy value retrieval with constant time complexity.

[0028] Step S5: At the second resolution level, for each configuration in the first selected configuration subset, call the first-level structure feature vector stored in the first-level cache unit to initialize the fine potential energy function calculation process. The fine potential energy function includes all-atom force field parameters, long-range Coulomb interaction terms, and implicit solvent model. The all-atomic force field parameters are obtained using a modified parameter set calibrated for the solvent environment. The long-range Coulomb interaction term is processed using the particle grid Ewald summation algorithm. The implicit solvent model is the generalized Born model, and the solvent dielectric constant is set to the standard value for the water environment.

[0029] Step S6: Perform fine potential energy function calculation to obtain the second-level high-precision energy value and energy gradient vector; Specifically: restore the original atomic-level coordinates of the configuration, load the full atomic force field parameter library, which covers the AMBER, CHARMM and OPLS series force fields; use the conjugate gradient algorithm to iteratively optimize the atomic positions until the energy gradient norm converges to below 0.001 electron volts per angstrom; and record the atomic coordinates, energy values ​​and gradient vectors simultaneously during the optimization process.

[0030] Step S7: Based on the second-level high-precision energy value, set a second energy threshold, remove configurations with energy values ​​higher than the second energy threshold, and retain the remaining configurations to form a second set of filtered configurations; Step S8: For each configuration in the second set of selected configurations, extract its full atomic coordinate offset, bond angle distribution histogram and solvent accessible surface area change rate to generate a second-level structure feature vector, and store the second-level structure feature vector, high-precision energy value and energy gradient vector together in the second-level cache unit. Specifically, the generation of the second-level structure feature vector includes the following steps: calculating the Euclidean norm of the atomic coordinate offsets before and after optimization; statistically analyzing the bond angle distribution histogram, dividing it into intervals of ten degrees; calculating the absolute value of the rate of change of solvent-accessible surface area; and combining the Euclidean norm, the bond angle distribution histogram, and the absolute value of the rate of change of solvent-accessible surface area into the second-level structure feature vector. The second-level cache unit uses a key-value pair database. The key consists of a low-dimensional code generated by principal component analysis after dimensionality reduction of the second-level structural feature vector. The value contains a high-precision energy value, an energy gradient vector, and a pointer to the original configuration identifier, supporting energy prediction of nearest neighbor configurations based on similarity metrics.

[0031] Step S9: At the third resolution level, for each configuration in the second selection configuration subset, call the second-level structure feature vector and energy gradient vector stored in the second-level cache unit as initial conditions input to the quantum mechanical semi-empirical method calculation module. The quantum mechanical semi-empirical method uses parameterized Hamiltonian to approximate the electronic structure. Its parameterized Hamiltonian is based on a pre-trained parameter library of specific element combinations. It uses high-precision fitting parameters for carbon, hydrogen, oxygen, and nitrogen atoms, medium-precision fitting parameters for sulfur and phosphorus atoms, and empirically shielded charge approximation for metal ions, ensuring the physical rationality of the electronic structure description under limited computing resources.

[0032] Step S10: Perform quantum mechanical semi-empirical method calculations to obtain the third-level electron energy correction value and orbital occupancy distribution, and store the above data in the energy correction cache unit; Specifically: Based on the gradient sensitivity analysis results in the second-level buffer, quantum correction is only performed on atoms with gradient magnitudes greater than 0.5 electron volts per angstrom and their neighboring atoms within the third order; the classical force field energy is used in the remaining regions; the total energy after correction is composed of the superposition of the quantum region energy, the non-quantum region energy, and the buffer layer coupling energy. The coupling energy of the buffer layer can be smoothly transitioned through a linear interpolation function, ensuring that the energy function is continuously differentiable throughout the entire domain. The interpolation weight of the linear interpolation function is determined by the distance of the atom from the boundary of the quantum region. The quantum mechanical semi-empirical method computation module adopts a block diagonalization strategy, which divides the molecular system into multiple subsystems according to the chemical bonding connectivity. The electronic Hamiltonian matrix is ​​solved independently for each subsystem, and the coupling between subsystems is corrected by the overlap integral of boundary atomic orbitals, which reduces the matrix solution dimension and improves computational efficiency.

[0033] Step S11: Superimpose the third-level electron energy correction value onto the second-level high-precision energy value to generate the final comprehensive energy assessment result; Step S12: Based on the final comprehensive energy assessment results, set the final energy ranking threshold, output the configurations with energy values ​​lower than the final energy ranking threshold as the preferred configuration set that meets the physical rationality standard and has controllable computational cost, and simultaneously output the energy value evolution curve, structural feature vector evolution trajectory and cache hit status log of each configuration at each resolution level for subsequent analysis and model optimization.

[0034] In addition, this application also provides a cost-controlled staged resolution generation and cache reuse system, specifically applied to the aforementioned cost-controlled staged resolution generation and cache reuse method, including the following modules: Configuration input interface module, used to receive the initial set of molecular configurations to be evaluated (i.e., step S1). The first resolution-level calculation module is used to calculate the coarse-grained energy value of each configuration in the initial molecular configuration set using a simplified potential energy function (i.e., step S2). The first-level filtering module is used to filter out configurations that meet the requirements and form a first set of filtered configurations (i.e., step S3). The first-level feature extraction and caching module is used to generate the first-level structural feature vector and associate the first-level structural feature vector with the corresponding coarse-grained energy value and store it in the first-level cache unit (i.e., step S4). The second resolution level calculation module is used to initialize the fine potential energy function calculation process and obtain the second-level high-precision energy value and energy gradient vector (i.e., steps S5 and S6). The second-level filtering module is used to filter out configurations that meet the requirements and form a second set of filtered configurations (i.e., step S7). The second-level feature extraction and caching module is used to generate the second-level structural feature vector and associate and store the second-level structural feature vector, high-precision energy value and energy gradient vector in the second-level cache unit (i.e. step S8). The third resolution level calculation module is used to obtain the third-level electron energy correction value and orbital occupancy distribution, and to superimpose the third-level electron energy correction value onto the second-level high-precision energy value to generate the final comprehensive energy assessment result (i.e., steps S9, S10 and S11). The final configuration output module is used to output configurations with energy values ​​lower than the final energy ranking threshold as a set of preferred configurations that meet physical rationality and have controllable computational costs (i.e., step S12).

[0035] The system also includes a resolution transition determination module, which is integrated between the second-level resolution calculation module and the third-level resolution calculation module. Its inputs are the energy gradient magnitude, bond angle standard deviation, and absolute value of the solvent-accessible surface area change rate calculated at the second level for the current configuration. Its output is a Boolean decision signal. When the energy gradient magnitude is less than the preset convergence threshold, the bond angle standard deviation is less than the structural stability threshold, and the absolute value of the solvent-accessible surface area change rate is less than the environmental disturbance tolerance threshold, it is determined that the configuration does not need to enter the third-level resolution and the high-precision energy value of the second level is directly used as the final evaluation result.

[0036] In the above method, the first-level cache unit, the second-level cache unit, and the energy correction cache unit together constitute a cross-level cache system. The first-level cache unit also stores the dimensionality reduction coordinates and statistical features of the first-level output and uses a Bloom filter to accelerate similarity retrieval. The second-level cache unit manages the optimized path data of the second level in the form of a circular queue and supports bidirectional traversal by timestamp and energy sorting. The energy correction cache unit adopts a B-tree index structure, which facilitates the quick location of historical correction records by energy range. Each cache unit is associated with a unified configuration hash code to ensure data consistency. When a new configuration enters any level, its hash code is calculated and used as the key to perform a cache matching query. If the match is successful, the cached data is retrieved and its timeliness is verified. If the timeliness exceeds the preset aging threshold, the cache refresh mechanism is triggered, the corresponding level calculation is re-executed and the cache content is updated. This mechanism ensures the freshness of the data and avoids the accumulation of invalid cache due to long-term inaccessibility.

[0037] To support large-scale parallel computing, this method designs a distributed cache synchronization protocol, enabling distributed deployment of cross-level caching systems. In a cluster environment, each compute node maintains a local cache copy and broadcasts new or updated cache entry metadata via a lightweight message bus. The broadcast content only includes configuration hash codes and metadata digests, without transmitting complete data blocks, thus reducing network overhead. When a node needs to access remote cached data, it first sends a query request to the coordinating node. The coordinating node locates the data holder based on the hash code and instructs it to directly push the complete data block to the requesting node. This decentralized architecture avoids single points of failure and ensures cache consistency through a version number mechanism: each cache entry is accompanied by an incrementing version number. When a node receives remote data, it compares the local version and only performs an update if the remote version is higher. Experiments show that in a thousand-core cluster, this protocol can control cache synchronization latency to the millisecond level, with negligible impact on overall computational efficiency.

[0038] During the multi-level collaborative evolution process, the system dynamically monitors the computational load and cache hit rate of each level and adjusts the transition judgment threshold accordingly. For example, when the average convergence step count of the second level consistently exceeds fifty steps, the energy screening threshold in the first resolution level is automatically relaxed, allowing more configurations to enter the fine calculation stage to distribute the computational pressure. Conversely, when the frequency of quantum computing requests in the third resolution level is too high, the activation conditions of the second level are tightened, retaining only the top 10% of configurations in terms of energy for quantum correction. This feedback adjustment mechanism ensures that the system can maintain optimal throughput under different hardware resource configurations and task scales. In addition, data integrity checks are performed before each level transition, including coordinate range checks, bond length rationality verification, and charge conservation tests. Failure of any check results in the configuration being marked as abnormal and subsequent computation being terminated, preventing erroneous data from polluting the cache or misleading decisions.

[0039] In practical deployment, the system provides a configuration interface for users to customize parameters at each level. Users can specify key parameters such as coarse-grained grid size, energy screening threshold, quantum correction activation conditions, and cache aging period, and can also upload custom force field files or semi-empirical method parameter sets. During the initialization phase, the system loads the user configuration and constructs the corresponding computational pipeline and caching strategy accordingly. For users lacking prior knowledge, the system has a built-in default parameter set, optimized based on benchmark test results of typical molecular systems such as protein folding, small organic molecule docking, and metal alloy phase transitions, covering most common application scenarios. During system operation, the real-time log module records the number of processed configurations at each level, average computation time, cache hit count, and reasons for abnormal termination, allowing users to analyze performance bottlenecks afterward. Log data is stored in structured text format and supports filtering and querying by time range, configuration type, or error code.

[0040] The core of this application lies in establishing a three-level progressive resolution architecture. Each level corresponds to different physical model precision and computational complexity. A cross-level cache matching mechanism enables the transfer and reuse of intermediate results. The first resolution level uses a coarse-grained model to quickly screen a massive number of initial configurations, retaining only a subset with potential value for the next stage. The second resolution level uses a full-atom force field to perform energy and gradient calculations on the screened configurations, while caching the calculated local structural features and gradient information to a dedicated storage unit. The third resolution level uses a quantum semi-empirical method to correct the energy for key configurations and determines whether to skip some redundant calculations based on the gradient distribution in the cache. Throughout the entire process, the output of each resolution level serves not only as input to the next stage but also as a benchmark for cache matching, identifying replaceable historical configurations in subsequent calculations and thus avoiding repeated high-cost computations.

[0041] In the first resolution level, the system first receives an initial set of configurations provided by an external input module. This set contains the spatial coordinates and topological connections of all molecular or material structures to be evaluated. Spatial dimensionality reduction is performed on each configuration, mapping its atomic coordinates to a pre-defined coarse-grained grid system. The grid spacing is pre-set based on the maximum tolerance error of the target system, typically ranging from 0.5 nm to 1.5 nm. After mapping, a simplified potential energy function is called to calculate the single-point energy of the coarse-grained configuration. The calculated energy value and the configuration's eigenvector in the coarse-grained space together constitute the preliminary screening criteria. The eigenvector consists of the local density, average coordination number, and spatial distribution entropy of each coarse-grained unit within the configuration. Energy thresholds and eigenvalue similarity thresholds are set. Only when the configuration's energy is below the threshold and its eigenvector has a cosine similarity to the cached configurations less than 0.8 is the configuration allowed to enter the second resolution level. This process effectively filters out more than 80% of low-potential configurations, significantly reducing the subsequent computational load.

[0042] Upon entering the second resolution level, the system restores the original atomic-level coordinates of the preserved configurations and loads a full-atom force field parameter library. This library covers standard force field descriptions of common organic molecules, inorganic crystals, and biomacromolecules, including but not limited to the AMBER, CHARMM, and OPLS series. Energy minimization calculations are performed sequentially for each configuration, using a conjugate gradient algorithm to iteratively optimize atomic positions until the energy gradient norm converges to below 0.001 electron volts per angstrom. During optimization, the atomic coordinates, energy values, and gradient vectors of each iteration are recorded synchronously. This data is categorized and stored in gradient cache units according to configuration identifiers. The gradient cache units are organized using a key-value pair structure, where the key is a unique hash code for the configuration, and the value is a composite data block containing the coordinate sequence, energy trajectory, and gradient tensor. To improve cache hit rate, before calculating each new configuration, its coarse-grained feature vector is extracted, and historical configurations with the same feature pattern are searched in the cache. If a match is found, its cached gradient is directly used as the initial search direction, thereby accelerating the convergence process of the current configuration. Experiments show that this strategy can reduce the average number of convergence steps by 30% to 50%, and is particularly effective when dealing with homologous series or molecular clusters with similar conformations.

[0043] After the configuration passes the second-level energy calculation, the system initiates the third-resolution level quantum semi-empirical correction process. This level is activated only for configurations that meet specific conditions, including: the final energy of the configuration in the second level is within 5% of the global minimum energy, or there are more than three local minimum oscillation records in its gradient buffer. Configurations that meet the conditions are sent to the semi-empirical quantum mechanical engine, which calculates its single-point total energy using the PM6 or AM1 method and outputs the electron density distribution and frontier orbital energy levels. To control computational costs, it is possible to choose not to perform full quantum calculations on all atoms, but instead, based on the gradient sensitivity analysis results in the second-level buffer, only perform quantum corrections on atoms with gradient amplitudes greater than 0.5 electron volts per angstrom and their neighboring atoms within three orders of energy. The remaining regions still use the classical force field energy. The corrected total energy consists of three superimposed parts: the semi-empirical energy of the quantum region, the classical force field energy of the non-quantum region, and the buffer layer coupling energy at the boundary between the two regions. The buffer layer coupling energy is smoothly transitioned through a linear interpolation function to ensure that the energy function is continuously differentiable throughout the entire domain. The difference between the final corrected energy and the original classical energy is recorded as the correction offset, and it is written into the energy decision cache along with the configuration identifier. If a similar configuration enters this level later, the energy decision cache is queried first. If the offset difference is less than 0.05 electron volts, the historical correction value is directly reused, and the quantum computing step is skipped.

[0044] The effectiveness of this method has been verified in several real-world cases, including: In a virtual screening task involving 100,000 candidate drug molecules, the traditional all-quantum computing scheme is expected to consume about 8 million core hours. However, by adopting this method, the total computing cost is reduced to 1.2 million core hours, a reduction of 85%, and the overlap rate between the top 100 molecules finally screened and the all-quantum benchmark results reaches 92%. In another structural optimization task for metal-organic frameworks, this method successfully identified three previously unreported low-energy stable configurations with formation energies 0.3 to 0.5 electron volts lower than those in the literature, demonstrating that the phased strategy did not sacrifice physical rationality. The cache reuse mechanism performed particularly well in this task: the gradient cache hit rate at the second level reached 68%, and the energy correction reuse rate at the third level also reached 41%, significantly reducing redundant computational overhead.

[0045] The above description is merely a preferred embodiment of the present invention; it encompasses all the protection scope of the present invention. Any equivalent substitutions or modifications made by those skilled in the art within the technical scope disclosed in the present invention, based on the technical solutions and improved concepts of the present invention, should be covered within the protection scope of the present invention.

Claims

1. A cost-controlled, phased resolution generation and cache reuse method, characterized by: Includes the following steps: Step S1: Receive the initial molecular configuration set to be evaluated through the configuration input interface module. The initial molecular configuration set contains multiple three-dimensional spatial coordinate sequences to be screened. Step S2: At the first resolution level, for each configuration in the initial molecular configuration set, a simplified potential energy function is used to calculate its coarse-grained energy value. The simplified potential energy function includes only bond length, bond angle and van der Waals repulsion terms. Step S3: Based on the coarse-grained energy value, set a first energy threshold, remove configurations with energy values ​​higher than the first energy threshold, and retain the remaining configurations to form a first set of filtered configurations; Step S4: For each configuration in the first set of selected configurations, extract its atomic arrangement topological features and local geometric invariants to generate a first-level structural feature vector, and associate the first-level structural feature vector with the corresponding coarse-grained energy value and store it in the first-level cache unit. Step S5: At the second resolution level, for each configuration in the first selected configuration subset, call the first-level structural feature vector stored in the first-level cache unit to initialize the fine potential energy function calculation process. The fine potential energy function includes all-atom force field parameters, long-range Coulomb interaction terms, and implicit solvent model. Step S6: Perform the fine potential energy function calculation to obtain the second-level high-precision energy value and energy gradient vector; Step S7: Based on the second-level high-precision energy value, set a second energy threshold, remove configurations with energy values ​​higher than the second energy threshold, and retain the remaining configurations to form a second set of filtered configurations; Step S8: For each configuration in the second set of selected configurations, extract its full atomic coordinate offset, bond angle distribution histogram and solvent accessible surface area change rate to generate a second-level structure feature vector, and store the second-level structure feature vector, high-precision energy value and energy gradient vector in the second-level cache unit. Step S9: At the third resolution level, for each configuration in the second selection configuration subset, call the second-level structure feature vector and energy gradient vector stored in the second-level cache unit as initial conditions input to the quantum mechanical semi-empirical method calculation module. The quantum mechanical semi-empirical method uses parameterized Hamiltonian to approximate the electronic structure. Step S10: Perform the quantum mechanical semi-empirical method calculation to obtain the third-level electron energy correction value and orbital occupancy distribution; Step S11: Superimpose the third-level electron energy correction value onto the second-level high-precision energy value to generate the final comprehensive energy assessment result; Step S12: Based on the final comprehensive energy assessment result, set a final energy ranking threshold, and output configurations with energy values ​​lower than the final energy ranking threshold as a set of preferred configurations that meet physical rationality standards and have controllable computational costs.

2. The cost-controlled phased resolution generation and cache reuse method according to claim 1, characterized in that: The specific operations of step S2 include the following: mapping the three-dimensional spatial coordinate sequence to a preset coarse-grained grid system, with the grid spacing ranging from 0.5 nanometers to 1.5 nanometers; performing single-point energy calculation on the mapped configuration, wherein the bond length term in the simplified potential energy function adopts the harmonic oscillator potential model, the bond angle term adopts the cosine expansion potential model, and the van der Waals repulsion term adopts the twelfth power inverse proportional potential model; all model parameters are extracted and fixed from the general biomolecular force field parameter set, and no dynamic fitting is performed.

3. The cost-controlled phased resolution generation and cache reuse method according to claim 2, characterized in that: The generation operation of the first-level structural feature vector in step S4 includes the following: calculating the local density, average coordination number and spatial distribution entropy of each coarse-grained unit in the configuration, and combining the local density, average coordination number and spatial distribution entropy into the first-level structural feature vector, which is used for subsequent similarity matching and cache retrieval.

4. The cost-controlled phased resolution generation and cache reuse method according to claim 3, characterized in that: The first-level cache unit adopts a hash table data structure, using a normalized floating-point array of the feature vectors of the first-level structure as the key and a coarse-grained energy value as the storage value, supporting feature matching and energy value retrieval with constant time complexity.

5. The cost-controlled staged resolution generation and cache reuse method according to claim 4, characterized in that: The specific operations of step S6 include the following: restoring the original atomic-level coordinates of the configuration, loading the full atomic force field parameter library, which covers the AMBER, CHARMM and OPLS series force fields; using the conjugate gradient algorithm to iteratively optimize the atomic positions until the energy gradient norm converges to below 0.001 electron volts per angstrom; and synchronously recording the atomic coordinates, energy values ​​and gradient vectors during the optimization process.

6. The cost-controlled staged resolution generation and cache reuse method according to claim 5, characterized in that: The generation operation of the second-level structure feature vector in step S8 includes the following: calculating the Euclidean norm of the atomic coordinate offset before and after optimization; statistically analyzing the bond angle distribution histogram, dividing it into intervals of ten degrees; calculating the absolute value of the rate of change of solvent-accessible surface area; and combining the Euclidean norm, the bond angle distribution histogram, and the absolute value of the rate of change of solvent-accessible surface area into the second-level structure feature vector.

7. The cost-controlled staged resolution generation and cache reuse method according to claim 6, characterized in that: The second-level cache unit uses a key-value pair database. The key consists of a low-dimensional code generated by principal component analysis after dimensionality reduction of the second-level structural feature vector. The value includes a high-precision energy value, an energy gradient vector, and a pointer to the original configuration identifier, supporting energy prediction of nearest neighbor configurations based on similarity metrics.

8. The cost-controlled staged resolution generation and cache reuse method according to claim 7, characterized in that: The specific operation of step S10 includes the following: based on the gradient sensitivity analysis results in the second-level buffer, quantum correction is performed only on atoms with gradient amplitudes greater than 0.5 electron volts per angstrom and their neighboring atoms within the third order; the classical force field energy is used in the remaining regions; the total energy after correction is composed of the superposition of the quantum region energy, the non-quantum region energy and the buffer layer coupling energy.

9. The cost-controlled staged resolution generation and cache reuse method according to claim 8, characterized in that: The coupling energy of the buffer layer can be smoothly transitioned through a linear interpolation function, ensuring that the energy function is continuously differentiable throughout the entire domain. The interpolation weight of the linear interpolation function is determined by the distance of the atom from the boundary of the quantum region.

10. The cost-controlled staged resolution generation and cache reuse method according to claim 9, characterized in that: The quantum mechanical semi-empirical method calculation module adopts a block diagonalization strategy, which divides the molecular system into multiple subsystems according to the chemical bond connectivity. The electronic Hamiltonian matrix is ​​solved independently for each subsystem, and the coupling between subsystems is corrected by the boundary atomic orbital overlap integral, thereby reducing the matrix solution dimension and improving the calculation efficiency.

11. The cost-controlled staged resolution generation and cache reuse method according to claim 10, characterized in that: The configuration input interface module supports parsing standard molecular file formats, including but not limited to three-dimensional coordinate files, connectivity description files, and charge distribution files. During import, it automatically performs atom type mapping, missing hydrogen atom completion, and initial geometry optimization to ensure that the input configuration meets the format and physical consistency requirements of subsequent calculation modules.

12. A cost-controlled staged resolution generation and cache reuse system, applied to the cost-controlled staged resolution generation and cache reuse method of claim 1, characterized in that: Includes the following modules: Configuration input interface module, used to receive the initial set of molecular configurations to be evaluated; The first resolution-level calculation module is used to calculate the coarse-grained energy value of each configuration in the initial molecular configuration set using a simplified potential energy function; The first-level filtering module is used to filter configurations that meet the requirements and form a first set of filtered configurations. The first-level feature extraction and caching module is used to generate the first-level structural feature vector and associate the first-level structural feature vector with the corresponding coarse-grained energy value and store it in the first-level cache unit. The second resolution level calculation module is used to obtain the high-precision energy value and energy gradient vector of the second level; The second-level filtering module is used to filter configurations that meet the requirements and form a second set of filtered configurations. The second-level feature extraction and caching module is used to generate the second-level structural feature vector and associate and store the second-level structural feature vector, high-precision energy value and energy gradient vector in the second-level cache unit. The third-resolution level calculation module is used to obtain the third-level electron energy correction value and orbital occupancy distribution, and generate the final comprehensive energy assessment result. The final configuration output module is used to output configurations with energy values ​​lower than the final energy ranking threshold as a set of preferred configurations that meet physical rationality standards and have controllable computational costs.

13. The cost-controlled staged resolution generation and cache reuse system according to claim 12, characterized in that: It also includes a resolution transition determination module, which is integrated between the second resolution level calculation module and the third resolution level calculation module. Its inputs are the energy gradient magnitude, bond angle standard deviation, and absolute value of the solvent-accessible surface area change rate calculated at the second level for the current configuration. Its output is a Boolean decision signal. When the energy gradient magnitude is less than a preset convergence threshold, the bond angle standard deviation is less than a structural stability threshold, and the absolute value of the solvent-accessible surface area change rate is less than an environmental disturbance tolerance threshold, it is determined that the configuration does not need to enter the third resolution level, and the high-precision energy value of the second level is directly used as the final evaluation result.

Images