An artificial intelligence application automatic development method and system for ocean science

Abstract

The application discloses an artificial intelligence application automatic development method and system for marine science, and the method is developed around three main lines: the first main line is a bottom architecture design, a modular and verifiable development pipeline is constructed through a pipeline-filter architecture style and a finite state machine theory; the second main line is a whole-link quality guarantee and an intelligent enhancement mechanism, along an entrance-process-outlet logic main line, three defense lines of resource access screening, intelligent guidance recommendation and automatic quality verification are respectively established; the third main line is to fuse the first two main lines into an end-to-end executable method. The application greatly reduces the technical threshold of marine science artificial intelligence development, enables non-computer background scientific researchers to release the core energy from engineering labor, and refocuses on scientific exploration itself.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an automated development method and system for artificial intelligence applications in marine science. Background Technology

[0002] Currently, there are multiple areas of research and practice both domestically and internationally related to artificial intelligence application development platforms, intelligent assistance, and quality assurance technologies: (1) In the field of general artificial intelligence, the industry has established relatively mature machine learning operations (MLOps) standards. Represented by Google's TFX (TensorFlow Extended), this platform defines a set of standardized component specifications, covering the entire pipeline from data ingestion, feature engineering, model training to service deployment. AutoKeras uses neural architecture search technology to automatically search for the optimal network structure, reducing the threshold for model selection. However, although such general tools are powerful, they lack native support for multidimensional scientific data (such as high-dimensional tensors and latitude and longitude grids), are cumbersome to configure, and have many parameters, which poses a huge cognitive burden for scientists without a computer science background.

[0003] (2) For a dedicated development ecosystem for Earth and Ocean sciences, the cloud-native Earth science analysis architecture proposed by the Pangeo community uses Xarray as the core data structure and combines it with the Dask distributed computing framework to successfully achieve interactive analysis of meteorological data. The OceanAI platform proposed by Chen et al. integrates authoritative data services from the National Oceanic and Atmospheric Administration (NOAA) and achieves real-time data retrieval and visualization chart generation driven by natural language through parameterized function calls and thought chain reasoning. However, the Pangeo community is essentially a data analysis tool and lacks support for model training, testing, and deployment processes; although OceanAI lowers the interaction threshold, it is still essentially a data query assistant and lacks many necessary capabilities for application development. Existing tools have failed to deeply decouple and automate the flow of tasks, data, models, and deployment.

[0004] (3) Vertical Domain Large Models and Retrieval Enhancement Generation Techniques. RAG technology, by introducing external knowledge bases and integrating retrieval and generation capabilities, has become a key technical route for breaking through knowledge barriers in vertical domains. Early dense paragraph retrieval used a dual-tower model to achieve semantic matching, but it was difficult to capture precise scientific terms. Recent research shows that a hybrid retrieval architecture combining sparse and dense vectors can simultaneously take into account both precise keyword matching and semantic generalization understanding. However, current RAG technologies mostly focus on static document knowledge extraction, with few studies involving recommendations for executable entities such as code entities and model weights, and generally lacking context awareness of the developer's current project stage.

[0005] (4) Intelligent Testing and Quality Assurance of AI Systems. In the field of automated testing, the ChatUniTest framework introduces an adaptive focus context mechanism to construct a closed-loop mechanism of generation-verification-repair. In terms of static code analysis and repair, LLM is used as a semantic-level filtering mechanism to remove invalid vulnerability warnings. However, existing testing tools mostly focus on the syntax correctness or routine logic checks at the code level, and are extremely lacking in dynamic defense mechanisms against runtime robustness boundaries in scientific computing (such as hidden memory leaks, extreme numerical overflows, and other abnormal states). In addition, there is currently no objective evaluation standard in academia that can comprehensively quantify the static code specifications, dynamic robustness of the model, and domain adaptability.

[0006] In summary, while existing platforms have made significant progress in their respective areas of expertise, no single tool can cover the entire chain of needs for marine science AI development: it lacks a reliable software supply chain mechanism for the marine field, context-aware intelligent guidance capabilities, and an export quality assurance system that integrates static and dynamic testing. Therefore, this invention proposes Galene, an automated development platform for AI applications in marine science. By introducing RAG and LLM technologies, it achieves full automation and intelligence from resource recommendation and application construction to quality verification. Summary of the Invention

[0007] The purpose of this invention is to provide an automated development method and system for artificial intelligence applications in marine science, which significantly lowers the technical threshold for the development of artificial intelligence in marine science, enabling researchers without a computer science background to free up their core energy from engineering work and refocus on scientific exploration itself.

[0008] To achieve the above objectives, the present invention provides the following solution: An automated development method for artificial intelligence applications in marine science includes the following steps: S1. Receive the marine artificial intelligence task type selected by the user and construct the marine artificial intelligence task into a task configuration data structure that includes data space constraint components, model architecture topology components, hyperparameter configuration space components, and optimization target definition components. S2. Based on a finite state machine, a development pipeline is built, which divides the development process into multiple execution stages. The process is driven by responding to input events through state transition functions. Before each state transition, a precondition check is performed, and the corresponding components in the task configuration data structure are instantiated and locked sequentially as the stage progresses. S3. Perform offline access screening on external data and model assets to be included in the development pipeline in order to establish a standardized data base and model matrix; S4. Construct a dual-path hybrid knowledge base that includes structured data channels and unstructured data channels, monitor the current development pipeline status and locked task parameters in real time, perform intent disambiguation and retrieval domain convergence on user queries based on the development pipeline status and context parameters, perform parallel retrieval and result fusion, and perform continuous experience evolution based on historical experimental data to provide intelligent recommendations. S5. Before the development pipeline moves to the delivery stage, static verification and dynamic detection and robust fuzz testing based on a large language model are performed on the generated standardized deployment package in sequence to calculate the comprehensive quality score. When a fatal defect is detected, a non-linear circuit breaker penalty mechanism is triggered to reduce the total score, thereby outputting a quality-verified deployment package.

[0009] Preferably, in S1, the task configuration data structure includes a data space constraint component, a model architecture topology component, a hyperparameter configuration space component, and an optimization objective definition component; wherein: The data space constraint component, including the verification rules for grid coordinates and sampling frequency, is mapped to data cleaning and tensor alignment, used to perform the projection mapping from the physical observation space to the computation tensor space; the model architecture topology component is transformed into a neural network computation graph instantiation script; the hyperparameter configuration space component is mapped to a valid parameter boundary matrix in memory, used to dynamically calculate the physical feasible domain of the control variables; the optimization objective definition component is compiled into executable loss function code logic; the four components together constitute a digital contract that is parsed and verified by the computer processor.

[0010] Preferably, in S2, the process is driven by responding to input events through a state transition function, and a precondition check is performed before each state transition, specifically including: When transitioning from the data preparation phase to the model selection phase, verify whether the computational layer dimension of the selected model is strictly isomorphic to the tensor declared by the data space constraint components. When transitioning from the parameter configuration phase to the training execution phase, it is verified whether the hyperparameter configuration is contained within the feasible region defined by the dynamically calculated legal parameter boundary matrix. The postconditions for each state transition include: the quaternion components of the previous stage have been instantiated, and the candidate set of the next stage has been automatically filtered to a subset that is compatible with the parameters locked in the previous stage.

[0011] Preferably, in S4, the dual-path hybrid knowledge foundation includes a relational database and a vector database, and the parallel retrieval and result fusion specifically includes a three-stage retrieval pipeline: The first stage is query understanding and instruction planning: the pre-cognitive agent classifies the user's query intent and links entities, and outputs a structured query plan containing structured query statements, vector search keywords and information priority indicators; The second stage is heterogeneous knowledge parallel retrieval: asynchronously and in parallel triggering the precise query channel of the relational database and the semantic retrieval channel of the vector database to obtain deterministic facts and principle knowledge fragments of historical optimal configuration respectively; The third stage is intent-based heuristic fusion: the token capacity quota of the two results in the context window is dynamically allocated according to the priority identifier. Factual queries are prioritized to fill the results of relational databases, methodological queries are prioritized to fill the results of vector databases, and recommendation decision queries are subject to balanced quota allocation between the two paths.

[0012] Preferably, in S4, the continuous experience evolution based on historical experimental data includes a three-layer progressive mechanism: Instance-level evolution: After each training session, record the experimental configuration and results, and update the globally optimal parameters for a specific data-model combination; Pattern-level evolution: When the number of successful samples reaches a preset threshold, it is automatically triggered. A robust statistical method based on interquartile range is used to extract high-confidence hyperparameter recommendation intervals from discrete successful experiment configurations, thus removing random noise interference from single experiments. Migration-level evolution: For a new dataset, it is mapped to a standardized feature vector containing four components: spatial resolution, temporal resolution, seasonality intensity, and data size. The cosine similarity between the new dataset and the existing dataset is calculated. When the cosine similarity exceeds a threshold, cross-domain experience copying is triggered.

[0013] Preferably, in S5, static verification includes three levels of defense executed sequentially: The first level of structural standardization verification checks whether the deployment package contains the core execution script and persistence weights; the second level of environment compatibility verification parses the dependency declaration file and detects version range mutual exclusion conflicts; the third level of code structure scanning based on abstract syntax trees verifies whether the model implements the standard interface for forward propagation and data iteration, calculates the cyclomatic complexity of the core functions, and issues refactoring warnings for high-complexity code blocks that exceed the threshold. The three-level verification follows the gating principle: the next level is only triggered after the previous level has passed.

[0014] Preferably, in S5, the formula for calculating the overall quality score is as follows:

[0015] in, These represent the normalized base scores for four dimensions: static standardization, code enhancement, dynamic detection, and robustness performance. For the corresponding weight coefficients This is a critical defect flag variable. When the sandbox detects a critical defect such as video memory overflow, process deadlock, or uncaught runtime crash, this variable is set to 1; otherwise, it is 0. The penalty factor is set to 0.4. Through this mechanism, once the model triggers a serious defect, the penalty factor will force a 40% reduction in the total score, achieving a veto-based bottom-line interception.

[0016] This invention also provides an automated development system for artificial intelligence applications in marine science, applied to an automated development method for artificial intelligence applications in marine science as described above, comprising: The user interaction layer is used to provide a visual development wizard interface, track user decision-making processes in real time, and maintain the global task context state. The core orchestration layer communicates with the user interaction layer, is configured to execute steps S1-S2 and S4, and manages the task configuration data structure as a task entity in memory. The infrastructure layer, which communicates with the core orchestration layer, is configured to execute S3 and S5 steps and provides data management units, model library management units, computation scheduling units, and code generation units.

[0017] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements an automated development method for artificial intelligence applications for marine science as described above.

[0018] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention constructs an automated development system based on a pipeline-filter architecture and finite state machines. It formalizes marine artificial intelligence tasks into a four-tuple data structure containing data space constraints, model architecture, hyperparameter space, and optimization objectives. This achieves end-to-end intelligent and low-code implementation from resource recommendation and application construction to quality verification, fundamentally eliminating the strong coupling between data processing and model training. This reduces cross-task switching time by 96.8%, end-to-end development time by 67.2%, and data adaptation code modification to zero lines, while maintaining statistically consistent model accuracy. Furthermore, it integrates a context-aware intelligent guidance system that combines relational and vector-based dual-path knowledge bases. The system improves the accuracy of fact recommendations to 86.3% and reduces the illusion rate to 10.0%, effectively overcoming the illusion and knowledge gap problems of large-scale models in vertical domains. Through a full-link quality governance system that combines entry screening, process guidance, and static three-level verification at the exit, as well as domain-aware fuzzy testing, and with the Qscore nonlinear circuit breaker penalty mechanism, it successfully implements a one-vote veto-based bottom-line interception for deployment packages with fatal defects such as memory overflow. Ultimately, it significantly lowers the technical threshold for the development of marine science artificial intelligence, enabling researchers without computer science backgrounds to free up their core energy from engineering work and refocus on scientific exploration itself. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A schematic diagram of the overall architecture of the automated development system for artificial intelligence applications in marine science provided by this invention; Figure 2 This is a schematic diagram of the three-order six-dimensional admission screening mechanism of the present invention; Figure 3 This is a diagram of the hybrid knowledge base architecture based on SQLite and ChromaDB in this invention. Figure 4 This is an overview diagram of the progressive structure and stability guarantee of the three-layer empirical evolution mechanism of the present invention; Figure 5 This is a schematic diagram of the step-by-step filtering structure of the three-level static verification defense line of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] This invention proposes an automated development method and system (Galene) for artificial intelligence applications in marine science. Its core design concept is to deeply embed best practices in software engineering into the scientific research workflow, enabling marine scientists without a computer science background to complete the entire process of developing artificial intelligence applications—from data preparation and model training to deployment—in a low-code manner. The platform's technical solution revolves around three main lines: the first is the underlying architecture design, which uses a pipeline-filter architecture style and finite state machine theory to build a modular and verifiable development pipeline; the second is the end-to-end quality assurance and intelligent enhancement mechanism, establishing three lines of defense along the entry-process-exit logic: resource access screening, intelligent guidance and recommendation, and automated quality verification; the third is the integration of the first two lines into an end-to-end executable system. The technical principles and implementation details of each part are described below.

[0024] This invention provides an automated development method for artificial intelligence applications in marine science, comprising the following steps: S1. Receive the marine artificial intelligence task type selected by the user and construct the marine artificial intelligence task into a task configuration data structure that includes data space constraint components, model architecture topology components, hyperparameter configuration space components, and optimization target definition components. S2. Based on a finite state machine, a development pipeline is built, which divides the development process into multiple execution stages. The process is driven by responding to input events through state transition functions. Before each state transition, a precondition check is performed, and the corresponding components in the task configuration data structure are instantiated and locked sequentially as the stage progresses. S3. Perform offline access screening on external data and model assets to be included in the development pipeline in order to establish a standardized data base and model matrix; S4. Construct a dual-path hybrid knowledge base that includes structured data channels and unstructured data channels, monitor the current development pipeline status and locked task parameters in real time, perform intent disambiguation and retrieval domain convergence on user queries based on the development pipeline status and context parameters, perform parallel retrieval and result fusion, and perform continuous experience evolution based on historical experimental data to provide intelligent recommendations. S5. Before the development pipeline moves to the delivery stage, static verification and dynamic detection and robust fuzz testing based on a large language model are performed on the generated standardized deployment package in sequence to calculate the comprehensive quality score. When a fatal defect is detected, a non-linear circuit breaker penalty mechanism is triggered to reduce the total score, thereby outputting a quality-verified deployment package.

[0025] Specifically, the method of the present invention includes: 1. Formal definition of task entities To unify the description of different types of marine artificial intelligence tasks such as time series prediction and water depth inversion, this invention will define any specific task T Defined as an ordered quadruple:

[0026] In the concrete implementation of a computer system, this formal model is transformed into a standardized task configuration data structure stored in a computer-readable storage medium. Specifically, the data space constraint components... This is visualized as a data cleaning and tensor alignment module within the system, used to perform projection mapping from the physical observation space to the computational tensor space, including verification rules for physical variables such as grid coordinates and sampling frequency; model architecture topology components. The script is transformed into a low-level neural network computation graph instantiation script, enabling the system to automatically adapt to the corresponding network topology; hyperparameter configuration spatial components. The physical boundary matrix is ​​mapped to the legal parameter boundaries in system memory. Based on this, the system dynamically calculates and constrains the physical feasible region of control variables such as learning rate and batch size; optimization target definition components. This is then compiled into executable loss function code logic. Through this formal model-data structure mapping, the four components together constitute a digital contract that can be parsed and verified by a computer processor, providing a solid foundation for the platform's automated code generation and static constraint verification.

[0027] 2. Pipeline lifecycle management based on finite state machines During user interaction, the components of the quadruple T need to be instantiated step by step according to specific rules. This invention formalizes the six-stage development process of the Pipeline into a deterministic finite state machine:

[0028] In the actual engineering deployment of the platform, this mathematical model is materialized as a pipelined state control engine driven by a computer processor. This involves a finite set of states. (Including the idle initial state) Core interactive states and final delivery state This is implemented in the global state registry in system memory as a series of discrete workflow state identifiers; input event set This corresponds to user input commands captured by the front-end interface and feedback signals from the back-end operation; state transition function. This is then compiled into an event routing script containing strict hard-gating logic. The transition rules for each stage and the instantiation effect on the four-tuple are shown in Table 1 (State Transition Rules of the Pipeline State Machine).

[0029] Table 1

[0030] In actual operation, the control engine listens to... Events in the process trigger the transition of the state pointer. Each execution... Before the state transition, the processor will call the underlying verification script to perform precondition comparisons (e.g., strong verification of whether the computational layer dimension of the selected model is consistent with the previous state). The declared tensor is strictly isomorphic, and if it fails, the transition is refused. This design successfully transforms the abstract state convergence theory into a system-level physical defense mechanism. As the state pointer moves forward, each component of the task quadruple is physically locked in sequence, thereby achieving early interception of configuration errors through the forced guidance of the computer program and supporting controlled step-by-step rollback trial and error.

[0031] 3. Entry Quality Assurance: Software Supply Chain Access Mechanism To address the inconsistent quality and security risks associated with open-source models, this invention establishes a three-tiered, six-dimensional offline admission screening standard. Each external asset proposed for inclusion in the platform must sequentially pass three review stages, such as... Figure 2 As shown.

[0032] Resources selected through an admission process are systematically reorganized according to the physical characteristics of the scientific mission. At the data level, the platform constructs a standardized data foundation covering multi-scale spatiotemporal features, encompassing long-term SST observation data from the Bohai Sea, Yellow Sea, East China Sea, and South China Sea, as well as high-resolution multispectral images of five typical islands and reefs, including the Maldives and the Qilianyu Islands. At the model level, the platform constructs a gradient model capability matrix, including nine heterogeneous architectures: for SST prediction tasks, it integrates high-precision Transformer variants such as PatchTST, iTransformer, and Informer, along with high-efficiency linear architectures such as DLinear and SOFTS; for depth inversion tasks, it integrates UNet++, Attention U-Net, and PhyUNet, which fuses physical information. This gradient design of the model matrix ensures that users can choose the most suitable solution based on their accuracy requirements and computational constraints.

[0033] 4. Intelligent Process Guidance: Retrieval and Recommendation Mechanism Based on Hybrid Knowledge Base Faced with complex hyperparameter configuration spaces, scientists without a computer science background often resort to inefficient, blind trial and error. This invention designs an intelligent guidance mechanism based on a hybrid knowledge base to provide users with accompanying parameter recommendations and fault diagnosis.

[0034] In terms of knowledge base architecture, this invention constructs a hybrid storage architecture based on the dual-path collaboration of a relational database (SQLite) and a vector database (ChromaDB). SQLite is responsible for storing static resource metadata (dataset attributes, model recommendation parameters) with strict structural constraints and dynamic experimental records, supporting precise conditional retrieval; ChromaDB is responsible for storing unstructured principle knowledge and domain rules (physical meaning of hyperparameters, fault diagnosis thought chains, functional boundary declarations, etc.), supporting semantic-level generalized reasoning. The two types of knowledge sources converge in the context window of the large language model at runtime through a specific fusion strategy, such as... Figure 3 As shown.

[0035] In terms of the retrieval pipeline, the system is designed with a three-stage processing flow. The first stage is query understanding and instruction planning: the pre-cognitive agent classifies user queries by intent and links entities, outputting a structured query plan containing SQL query statements, vector retrieval keywords, and information priority identifiers. The second stage is heterogeneous knowledge parallel retrieval: the SQL channel executes precise queries to obtain deterministic facts such as historical optimal configurations, while the Vector channel recalls Top-K principle knowledge fragments through cosine similarity. The two channels operate asynchronously in parallel to ensure real-time response. The third stage is intent-based heuristic fusion: the system dynamically allocates token capacity quotas for the two paths of results in the context window based on priority identifiers—factual queries prioritize filling SQL results, methodological queries prioritize filling the vector library's screening thought chain, and recommendation decision-making queries implement a balanced quota allocation between the two paths.

[0036] In terms of context awareness, this invention deeply couples the intelligent guidance mechanism with the pipeline state machine. The system maintains a dynamic session context load in the background, tracking key parameters such as the current pipeline state machine pointer, locked datasets, and model identifiers in real time, and implements precise convergence of the retrieval domain through two dimensions. The first dimension is contextual default completion based on entity parsing: when a user's question lacks key entities, the system automatically extracts locked identifiers from the session context and injects them into the retrieval conditions, eliminating the burden of the user repeatedly declaring background information; when the user explicitly specifies cross-domain entities, the system respects the user's intent and overrides the context default values. The second dimension is intent disambiguation based on state machine pointers: isomorphic natural language questions are mapped to completely different knowledge domains at different pipeline stages—for example, "How should the learning rate be adjusted?" is parsed as a numerical calibration requirement (SQL priority) in the parameter configuration stage and as a fault diagnosis requirement (Vector priority) in the training monitoring stage.

[0037] Regarding continuous experience evolution, this invention designs a three-tiered progressive evolution mechanism from instance level to pattern level and then to transfer level, enabling the knowledge base to continuously improve recommendation quality as usage increases, such as... Figure 4As shown. Instance-level evolution records the experimental configuration and results after each training iteration and updates the globally optimal parameters for a specific data-model combination. Pattern-level evolution is automatically triggered after the number of successful samples reaches a threshold. It employs a robust statistical method based on interquartile range (IQR) to extract high-confidence hyperparameter recommendation intervals from discrete successful experiments, effectively removing random noise interference from individual experiments. Transfer-level evolution addresses the cold start problem of entirely new datasets by mapping each dataset to a standardized feature vector containing four components: spatial resolution, temporal resolution, seasonality intensity, and data size. It calculates the cosine similarity between the new dataset and the existing dataset, triggering cross-domain experience copying when the similarity exceeds a threshold. This three-layer mechanism belongs to a parameter-free external knowledge accumulation paradigm. The experience accumulation process is fully interpretable and auditable, and immune to catastrophic forgetting problems.

[0038] 5. Export quality assurance: Layered verification system Once the user completes training, the platform automatically generates a standardized deployment package containing model weights, core source code, and environment setup scripts. This invention automatically performs a two-stage quality verification process, progressing from basic to advanced levels, before delivery.

[0039] The first phase is a three-tiered verification defense based on static analysis, such as... Figure 5 As shown. The first level is structural compliance verification, checking whether the delivery package contains key assets such as core execution scripts and persistent weights. The second level is environment compatibility verification, parsing dependency declaration files and detecting version range mutual exclusion conflicts. The third level is code structure scanning based on AST, verifying whether the model implements standard interfaces such as forward propagation and data iteration, calculating the cyclomatic complexity of core functions, and issuing refactoring warnings for high-complexity code blocks that exceed the threshold. The three-level verification follows a strict gating principle: the next level is only triggered after the previous level passes.

[0040] The second phase involves LLM-driven dynamic enhancement and detection. First, the system extracts AST features from the source code and inputs them into a large language model, automatically injecting structured document annotations into core modules. Second, the large model parses the code's runtime parameters, automatically synthesizing a detection instruction set covering baseline parameters and boundary extrema. Resource probes are attached to an isolated sandbox to perform dynamic inference; if continuous monotonic memory leakage or peak values ​​exceeding hardware limits are detected, the system directly intercepts the process. Finally, the large model, combined with the task context, generates edge anomaly scenarios consistent with marine physics, performing domain-aware robust fuzz testing, focusing on capturing whether unhandled runtime crashes occur and whether abnormal degradation branches are effectively triggered.

[0041] Finally, based on the above verification and testing results, the system calculates the overall quality score (Qscore) of the deployment package. The Qscore calculation includes a base score and a non-linear circuit breaker penalty mechanism, and its formula is as follows:

[0042] in, These represent the normalized base scores for four dimensions: static standardization, code enhancement, dynamic detection, and robustness performance. These are the corresponding weighting coefficients. This is a critical defect flag variable. When the sandbox detects a critical defect such as video memory overflow, process deadlock, or uncaught runtime crash, this variable is set to 1; otherwise, it is 0. The penalty factor is set to 0.4. Through this mechanism, once a model triggers a serious defect, the penalty factor will force a 40% reduction in the total score, thereby ensuring that any deployment package with a risk of runtime crash cannot obtain a qualified rating, achieving a veto-based bottom-line interception.

[0043] This invention also provides an automated development system for artificial intelligence applications in marine science, applied to an automated development method for artificial intelligence applications in marine science as described above, comprising: The user interaction layer is used to provide a visual development wizard interface, track user decision-making processes in real time, and maintain the global task context state. The core orchestration layer communicates with the user interaction layer, is configured to execute steps S1-S2 and S4, and manages the task configuration data structure as a task entity in memory. The infrastructure layer, which communicates with the core orchestration layer, is configured to execute S3 and S5 steps and provides data management units, model library management units, computation scheduling units, and code generation units.

[0044] Specifically, the overall framework and three-tier architecture model are as follows: This invention first addresses the practical pain points in the development of artificial intelligence in marine science, systematically summarizing five core engineering challenges, and then proposes seven core platform characteristics as a systematic response at the architectural level. The mapping relationship between the five types of challenges and the seven characteristics is shown in Table 2 (Core Challenges and Platform Characteristic Response Matrix).

[0045] Table 2

[0046] At the physical architecture level, the Galene platform adopts a strictly decoupled three-layer architecture model, orthogonally decomposing the complex task lifecycle into three independent responsibility domains. The user interaction layer, serving as the semantic translation interface between scientists and the computing system, uses a modern web graphical interface to visualize abstract workflows as linear guided steps, tracking user decision-making trajectories in real time and maintaining context state. The core orchestration layer, as the semantic hub of the system, undertakes three responsibilities: process management, constraint verification, and resource scheduling. It receives input events from the interaction layer, executes state transition functions, and verifies preconditions and postconditions. The infrastructure layer, as the physical foundation for platform operation, provides standardized resource services and an execution environment. Its internal data management and model library management units are responsible for implementing static access control, the computing scheduling unit is responsible for computing power allocation, and the code generation unit, relying on a metacode template engine, transforms instantiated task configurations into final code packages. The three-layer architecture and inter-layer data flow are as follows: Figure 1 As shown.

[0047] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements an automated development method for artificial intelligence applications for marine science as described above.

[0048] This invention deeply integrates best practices in software engineering into the artificial intelligence development workflow of marine science, achieving significant advantages over existing technologies in terms of architecture, intelligent assistance, and quality assurance.

[0049] At the architectural level, this invention adopts a pipeline-filter architecture style and finite state machine theory, deconstructing the originally highly coupled scientific research and development process into six loosely coupled stages, and uniformly describing different types of marine tasks using a four-tuple formal model. This design fundamentally eliminates the code entanglement between data processing and model training, allowing scientists to switch research tasks without rewriting the underlying code. Actual comparative experiments show that, compared with the traditional development model using standardized open-source algorithm libraries, this platform reduces the cross-task switching time from 69.5 minutes to 2.2 minutes, a reduction of 96.8%; end-to-end development time is shortened from 75.6 minutes to 24.8 minutes, a reduction of 67.2%; and the amount of code modification required for data adaptation is reduced from 19 lines to zero lines. At the same time, the model accuracy produced by the three development models under the same hyperparameter configurations does not show statistically significant differences, verifying that the platform's highly automated engineering encapsulation does not come at the expense of scientific computing accuracy.

[0050] At the intelligent assistance level, this invention effectively solves the illusion and knowledge gap problems of large language models in vertical domain applications through the collaborative design of a hybrid knowledge base and a context-aware retrieval pipeline. Experimental results show that the fact accuracy of the complete system reaches 86.3%, an improvement of 61.8 percentage points compared to the original large model without knowledge base access; the illusion rate is reduced from 43.3% to 10.0%. Ablation experiments further demonstrate that the hybrid architecture of SQL and vectors contributes 8.5 percentage points to the fact accuracy improvement compared to the pure vector scheme, and the context-aware mechanism contributes an additional 21.8 percentage points of gain on this basis. The three-layer continuous experience evolution mechanism enables the system's recommendation yield to show a continuous growth trend with the accumulation of experiments. In the cold start scenario, cross-domain migration based on physical similarity reduces the validation set loss of the first exploration by 46.8% compared to random search, effectively solving the problem of wasted computing power in the early stage of new dataset access.

[0051] At the quality assurance level, the export quality mechanism constructed in this invention, which combines static verification and dynamic fuzz testing, fills the gap in existing tools for robustness verification during the runtime of research code. In the full-process testing of nine positive models and one negative control model with injected defects integrated into the platform, AST scanning accurately located the structural complexity risks in the core algorithm, and domain fuzz testing successfully exposed the defensive programming defects of the physical constraint layer under abnormal inputs. Qscore's nonlinear circuit breaker mechanism forcibly suppressed the control model with injected memory leakage defects from a potentially acceptable level to 44.9 points, verifying the evaluation system's ability to intercept fatal defects. This mechanism ensures that all model assets flowing into the downstream production environment undergo systematic quality verification.

[0052] Regarding user acceptance, this platform has been deployed in the field by the East China Sea Bureau of the Ministry of Natural Resources. An evaluation experiment by eight marine scientists showed that the technology acceptance model questionnaire achieved high scores of 4.62, 4.84, and 4.75 (out of 5.00) in perceived usefulness, perceived ease of use, and future use intention, respectively. Among them, 87.5% of the participants only had junior or intermediate software engineering experience, indicating that this invention effectively lowers the technical threshold for interdisciplinary development, allowing scientists in the field to free up their core cognitive resources from engineering labor and refocus on scientific exploration itself.

[0053] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0054] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. An automated development method for artificial intelligence applications in marine science, characterized in that, Includes the following steps: S1. Receive the marine artificial intelligence task type selected by the user and construct the marine artificial intelligence task into a task configuration data structure that includes data space constraint components, model architecture topology components, hyperparameter configuration space components, and optimization target definition components. S2. A development pipeline is built based on a finite state machine. The development process is divided into multiple execution stages. The process is driven by responding to input events through state transition functions. Before each state transition, a precondition check is performed. As the stages progress, the corresponding components in the task configuration data structure are instantiated and locked in sequence. S3. Perform offline access screening on external data and model assets to be included in the development pipeline in order to establish a standardized data base and model matrix; S4. Construct a dual-path hybrid knowledge base that includes structured data channels and unstructured data channels, monitor the current development pipeline status and locked task parameters in real time, perform intent disambiguation and retrieval domain convergence on user queries based on the development pipeline status and context parameters, perform parallel retrieval and result fusion, and perform continuous experience evolution based on historical experimental data to provide intelligent recommendations. S5. Before the development pipeline moves to the delivery stage, static verification and dynamic detection and robust fuzz testing based on a large language model are performed on the generated standardized deployment package in sequence to calculate the comprehensive quality score. When a fatal defect is detected, a non-linear circuit breaker penalty mechanism is triggered to reduce the total score, thereby outputting a quality-verified deployment package.

2. The automatic development method for artificial intelligence applications in marine science according to claim 1, characterized in that, In S1, the task configuration data structure includes a data space constraint component, a model architecture topology component, a hyperparameter configuration space component, and an optimization objective definition component; wherein: The data space constraint component includes verification rules for grid coordinates and sampling frequency, which are mapped to data cleaning and tensor alignment for performing projection mapping from the physical observation space to the computation tensor space; the model architecture topology component is converted into a neural network computation graph instantiation script; the hyperparameter configuration space component is mapped to a valid parameter boundary matrix in memory for dynamically calculating the physical feasible domain of control variables; the optimization objective definition component is compiled into executable loss function code logic; the four components together constitute a digital contract that is parsed and verified by the computer processor.

3. The automatic development method for artificial intelligence applications in marine science according to claim 1, characterized in that, In step S2, the process is driven by responding to input events through a state transition function. Pre-condition checks are performed before each state transition, specifically including: When transitioning from the data preparation phase to the model selection phase, verify whether the computational layer dimension of the selected model is strictly isomorphic to the tensor declared by the data space constraint components. When transitioning from the parameter configuration phase to the training execution phase, it is verified whether the hyperparameter configuration is contained within the feasible region defined by the dynamically calculated legal parameter boundary matrix. The postconditions for each state transition include: the quaternion components of the previous stage have been instantiated, and the candidate set of the next stage has been automatically filtered to a subset that is compatible with the parameters locked in the previous stage.

4. The automatic development method for artificial intelligence applications in marine science according to claim 1, characterized in that, In S4, the dual-path hybrid knowledge foundation includes a relational database and a vector database, and the parallel retrieval and result fusion specifically includes a three-stage retrieval pipeline: The first stage is query understanding and instruction planning: the pre-cognitive agent classifies the user's query intent and links entities, and outputs a structured query plan containing structured query statements, vector retrieval keywords and information priority indicators; The second stage is heterogeneous knowledge parallel retrieval: asynchronously and in parallel triggering the precise query channel of the relational database and the semantic retrieval channel of the vector database to obtain deterministic facts and principle knowledge fragments of historical optimal configuration respectively; The third stage is intent-based heuristic fusion: the token capacity quota of the two results in the context window is dynamically allocated according to the priority identifier. Factual queries are prioritized to fill the relational database results, methodological queries are prioritized to fill the vector database results, and recommendation decision queries are subject to balanced quota allocation between the two paths.

5. The automatic development method for artificial intelligence applications in marine science according to claim 1, characterized in that, In S4, the continuous empirical evolution based on historical experimental data includes a three-layer progressive mechanism: Instance-level evolution: After each training session, record the experimental configuration and results, and update the globally optimal parameters for a specific data-model combination; Pattern-level evolution: When the number of successful samples reaches a preset threshold, it is automatically triggered. A robust statistical method based on interquartile range is used to extract high-confidence hyperparameter recommendation intervals from discrete successful experiment configurations, thus removing random noise interference from single experiments. Migration-level evolution: For a new dataset, it is mapped to a standardized feature vector containing four components: spatial resolution, temporal resolution, seasonality intensity, and data size. The cosine similarity between the new dataset and the existing dataset is calculated. When the cosine similarity exceeds a threshold, cross-domain experience copying is triggered.

6. The automatic development method for artificial intelligence applications in marine science according to claim 1, characterized in that, In S5, static verification includes three levels of defense executed sequentially: The first level of structural compliance verification checks whether the deployment package contains the core execution script and persistent weights; the second level of environment compatibility verification parses the dependency declaration file and detects version range mutual exclusion conflicts. The third level is a code structure scan based on an abstract syntax tree, which verifies whether the model implements the standard interface for forward propagation and data iteration, calculates the cyclomatic complexity of the core functions, and issues refactoring warnings for high-complexity code blocks that exceed the threshold. The three-level verification follows the gating principle: the next level is only triggered after the previous level has passed.

7. The automatic development method for artificial intelligence applications in marine science according to claim 1, characterized in that, In S5, the formula for calculating the overall quality score is as follows: in, These represent the normalized base scores for four dimensions: static standardization, code enhancement, dynamic detection, and robustness performance. For the corresponding weight coefficients This is a critical defect flag variable. When the sandbox detects a critical defect such as video memory overflow, process deadlock, or uncaught runtime crash, this variable is set to 1; otherwise, it is 0. The penalty factor is set to 0.

4. Through this mechanism, once the model triggers a serious defect, the penalty factor will force a 40% reduction in the total score, achieving a veto-based bottom-line interception.

8. An automated development system for artificial intelligence applications in marine science, applied to executing the automated development method for artificial intelligence applications in marine science as described in any one of claims 1-7, characterized in that, include: The user interaction layer is used to provide a visual development wizard interface, track user decision-making processes in real time, and maintain the global task context state. The core orchestration layer, which is connected to the user interaction layer, is configured to execute steps S1-S2 and S4, and manages the task configuration data structure as a task entity in memory. The infrastructure layer, which is communicatively connected to the core orchestration layer, is configured to execute steps S3 and S5 and provides a data management unit, a model library management unit, a computation scheduling unit, and a code generation unit.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements an automated development method for artificial intelligence applications for marine science as described in any one of claims 1-7.

Images