Method and system for real-time three-dimensional restoration of paleontology based on artificial intelligence and multi-modal interaction

Abstract

This invention relates to the field of artificial intelligence technology and discloses a method and system for real-time 3D paleontological reconstruction based on artificial intelligence and multimodal interaction. It aims to solve the problems of fragmented multi-source data, low automation, limited interaction methods, and lack of intelligent reasoning capabilities in existing technologies, which lead to strong subjectivity and difficulty in dynamic verification. The method includes: constructing a four-dimensional unified representation framework integrating fossil point clouds, stratigraphy, isotopes, and extant analog data, and generating a structured knowledge graph; using conditional generative adversarial networks to complete missing structures under biomechanical and evolutionary rules constraints, generating an anatomically consistent 3D morphological model; deploying natural language and virtual reality interfaces to achieve multimodal human-computer collaborative interaction; and triggering incremental learning based on user feedback and new evidence to complete real-time iterative optimization. This system integrates the above modules and scientific verification mechanisms, enabling high-fidelity, interactive, verifiable, and continuously evolving 3D dynamic paleontological reconstruction.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method and system for real-time three-dimensional reconstruction of paleontology based on artificial intelligence and multimodal interaction. Background Technology

[0002] With the deep integration of artificial intelligence and digital humanities technologies, paleontological research is gradually entering a new stage of intelligence and visualization. As an important carrier for recording the evolution of life on Earth, paleontological fossils have complex morphological structures and diverse information dimensions. Traditional restoration methods rely heavily on expert experience and two-dimensional image deduction, making it difficult to efficiently and accurately restore their three-dimensional dynamic characteristics and ecological behavior patterns.

[0003] Among them, paleontological reconstruction technology based on multimodal interaction aims to integrate fossil scanning data, geological environment information, biomechanical models, and motion data of modern analog species to achieve high-fidelity reconstruction of the morphology and behavior of extinct organisms through human-computer collaboration. The core objective of this technology direction is to overcome the limitations of static reconstruction and construct an interactive, verifiable, and evolving dynamic three-dimensional reconstruction system.

[0004] Current technologies for 3D paleontological reconstruction still suffer from significant shortcomings: First, data sources are fragmented; multi-source heterogeneous data such as fossil point clouds, sedimentary bedding, and isotope analysis lack a unified characterization framework, making effective alignment and fusion difficult. Second, the reconstruction process is highly dependent on human intervention, with low automation, hindering real-time iteration and dynamic correction. Third, interaction methods are limited; researchers cannot intuitively participate in reconstruction decisions using multimodal means such as natural language, gestures, or virtual reality. Finally, existing systems generally lack AI-based reasoning capabilities, failing to automatically fill in missing structures or simulate reasonable movement patterns based on limited fossil evidence, leading to highly subjective reconstruction results and difficulties in scientific verification. These problems severely restrict the in-depth application of paleontological reconstruction in scientific research, education, museum displays, and public communication, necessitating a new paradigm for real-time 3D reconstruction that integrates artificial intelligence and multimodal interaction. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for real-time three-dimensional reconstruction of paleontology based on artificial intelligence and multimodal interaction, which can effectively solve the problems in the background art mentioned above.

[0006] To achieve the above objectives, the technical solution adopted by this invention is as follows: a method and system for real-time 3D reconstruction of paleontology based on artificial intelligence and multimodal interaction, comprising the following specific steps: Step 1: Constructing a unified representation framework for multi-source heterogeneous data: collecting high-precision 3D point cloud data of paleontological fossils, sedimentary rock stratification information, isotopic dating data, and anatomical and kinematic parameters of modern closely related species; mapping the above multi-source heterogeneous data to a unified four-dimensional spatiotemporal coordinate system through a spatiotemporal alignment algorithm to form a structured paleontological knowledge graph; Step 2: Generating an initial 3D morphological model: based on the structured paleontological knowledge graph, using a conditional generative adversarial network to perform semantically perceptual geometric completion of missing parts of the fossil, wherein the generator takes the fossil point cloud as conditional input, and the discriminator integrates biomechanical constraints and evolutionary rationality rules to output a complete 3D morphological mesh with anatomical consistency; Step 3: Constructing a dynamic behavior simulation... The true engine imports the complete 3D morphological mesh into the physical simulation environment, combining the muscle attachment point inference module, joint degree of freedom constraint module, and gait pattern matching module to simulate the movement posture of paleontology under different terrains and speeds, generating dynamic behavior sequences that conform to biomechanical principles; Step four realizes multimodal human-computer collaborative interaction: deploying a natural language understanding interface, gesture recognition module, and virtual reality visualization interface, allowing researchers to adjust restoration parameters through voice commands, rotate or section the model through gesture operations, and verify the ecological rationality of the restoration results through immersive verification in a virtual reality environment; Step five executes a real-time iterative optimization mechanism: dynamically updating the structured paleontological knowledge graph based on user interaction feedback and new fossil evidence, and triggering the incremental learning process of the conditional generative adversarial network and dynamic behavior simulation engine, completing model correction and re-rendering within a preset time period, achieving closed-loop real-time restoration.

[0007] Preferably, in step one, the high-precision three-dimensional point cloud data is acquired by a laser confocal scanner, and the scanning resolution and point cloud density meet the requirements of high-fidelity modeling. The sedimentary rock bedding information is reconstructed by micro-CT tomography, and the layer thickness accuracy reaches the level required for geological analysis. The time error of the isotopic age data is controlled within the preset age error range after Bayesian correction.

[0008] Preferably, in step two, the generator of the conditional generative adversarial network adopts an encoder-decoder architecture. The encoder consists of multiple layers of three-dimensional convolutional layers, and the kernel size and number of channels of each layer are configured layer by layer according to the feature extraction requirements. The decoder fuses multi-scale features through skip connections and outputs vertex coordinates and normal vectors. The discriminator introduces a biomechanical loss function, which includes a bone strength constraint term, a muscle torque balance term, and an evolutionary distance penalty term. The weight coefficients of each term are dynamically set according to the restoration target.

[0009] Preferably, in step three, the muscle attachment point inference module uses a graph neural network based on a database of tendon distribution in modern birds and reptiles to predict the origin and insertion points of muscles in extinct species, achieving a prediction accuracy that meets the predetermined precision requirements; the joint degree of freedom constraint module calculates the maximum flexion and extension angles based on the curvature of fossil articular surfaces, with the error range controlled within a preset angle tolerance; and the gait pattern matching module retrieves the most similar gait templates from a motion database containing various extant tetrapods, with the matching similarity threshold set to a specific reference value.

[0010] Preferably, in step four, the natural language understanding interface supports the parsing of Chinese professional terms and can recognize multiple paleontology academic commands; the gesture recognition module captures key points of the hand based on a depth camera, and the recognition accuracy meets the requirements of interactive reliability, supporting four basic operations: zooming, panning, sectioning, and annotation; the refresh rate, field of view, and latency of the virtual reality visualization interface all meet the performance standards of immersive interaction.

[0011] Preferably, the incremental learning process in step five employs a knowledge distillation mechanism, injecting new evidence as soft labels into the original model, retaining historical restoration knowledge while absorbing new information, and controlling the amount of model parameter updates within a predetermined proportion of the total number of parameters to ensure the continuity and stability of the restoration results.

[0012] Preferably, the structured paleontological knowledge graph is stored using an attribute graph model. Node types include fossil specimens, stratigraphic units, isotope samples, and extant analogous species. Edge types include "excavated from", "belongs to", "anatomically homologous", and "motorally similar". The graph scale supports efficient querying of large-scale nodes and relationships.

[0013] Preferably, the dynamic behavior simulation engine integrates rigid body dynamics and soft tissue deformation models. The skeletal part adopts finite element analysis, and the mesh unit size is adapted to the details of biological structures. The soft tissue part adopts a mass-spring system, and the spring constant is dynamically adjusted according to the muscle type. The simulation frame rate is stabilized at the preset working frequency.

[0014] Preferably, the system also includes a scientific verification module, which automatically calculates the geometric overlap between the restoration model and the original fossil data, the energy efficiency index of the movement trajectory, and the ecological niche matching degree. All three indicators must be greater than or equal to a preset validity threshold to be considered a valid restoration.

[0015] Preferably, the method is deployed on a distributed computing platform, with the front-end interactive terminal connected to the cloud inference server via a high-speed wireless network. Model generation and rendering tasks are processed in parallel by a GPU cluster, and the time taken for a single complete restoration process is less than a preset time limit, supporting concurrent collaboration among multiple researchers.

[0016] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0017] 1. Deep fusion capability of multi-source data

[0018] Breaking through data silos: By constructing a unified four-dimensional spatiotemporal representation framework, the organic integration of fossil morphology, geological age, environmental background and extant analog data is achieved for the first time. Data alignment errors are controlled within the preset spatial tolerance, laying the foundation for high-fidelity reconstruction. Knowledge-driven reconstruction: The structured paleontological knowledge graph transforms expert experience into computable rules, significantly reducing the subjectivity and arbitrariness of the reconstruction process and improving scientific rigor.

[0019] 2. High-precision automated modeling capability

[0020] Intelligent geometric completion: Under the dual constraints of biomechanics and evolutionary rules, conditional generative adversarial networks significantly improve the accuracy of completing missing parts compared with traditional interpolation methods, especially suitable for highly variable regions such as the skull and extremities; Real-time dynamic simulation: Integrating a physics engine and a biomechanical model, it can generate complex behaviors such as walking, running, and predation that conform to the principle of energy optimization, breaking through the limitations of static restoration.

[0021] 3. Natural and efficient human-computer collaboration experience

[0022] Multimodal interactive intuitiveness: Researchers can directly control the restoration process through natural language and gestures without programming, which greatly reduces the cost of interactive learning and significantly improves decision-making efficiency; Immersive verification environment: The virtual reality interface supports seamless switching from macroscopic ecological scenes to microscopic anatomical structures, which facilitates multidimensional cross-validation of restoration hypotheses.

[0023] 4. Closed-loop continuous evolution mechanism

[0024] Real-time iterative optimization: The system can respond to new evidence or user feedback within a preset time period, realizing a closed loop of "hypothesis-verification-correction" in scientific research and significantly shortening the restoration cycle; Knowledge accumulation and sharing: All restoration versions and interaction records are automatically stored in the knowledge graph, forming traceable and reusable digital assets, promoting the digital transformation of paleontological research paradigms. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall technical solution of the method and system of the present invention;

[0026] Figure 2 This is a schematic diagram illustrating the core principle of generating the initial three-dimensional morphological model driven by the fusion of conditional generative adversarial networks and biomechanical constraints in this invention.

[0027] Figure 3 This is a schematic diagram of the logic flow of the dynamic behavior simulation engine in this invention;

[0028] Figure 4This is a schematic diagram of the multi-level interaction relationship and data flow of the multimodal human-computer collaborative interaction and real-time iterative optimization mechanism in this invention. Detailed Implementation

[0029] 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.

[0030] Example 1

[0031] This embodiment is applied to a digital reconstruction laboratory in a paleontology research institution, used for high-fidelity three-dimensional morphological reconstruction and dynamic behavior simulation of Late Jurassic theropod dinosaur fossil specimens. A complete "real-time three-dimensional paleontological reconstruction system based on artificial intelligence and multimodal interaction" is deployed in this scenario. Its hardware architecture consists of three parts: a front-end interactive terminal, edge computing nodes, and a cloud inference server, achieving low-latency collaboration through a 5G private network.

[0032] In terms of system architecture, the system includes a data acquisition subsystem, a knowledge graph construction and storage subsystem, an AI modeling and simulation subsystem, a multimodal interaction subsystem, and a scientific verification and iterative optimization subsystem. Each subsystem is interconnected via a high-speed internal bus and standardized API interfaces, forming a closed-loop data flow.

[0033] The data acquisition subsystem includes a laser confocal scanner (model: Leica TCS SP8 DIVE), equipped with a three-wavelength excitation source of 405nm / 488nm / 561nm and an objective lens numerical aperture of 1.4, capable of acquiring point clouds of fossil surfaces with a spatial resolution of 10 micrometers, and maintaining a point cloud density stably above 520 points / square millimeter; a micro-CT tomography scanner (model: Zeiss Xradia 520 Versa), with an X-ray source voltage range of 20–160 kV and a detector pixel size of 7.4 μm, used to reconstruct the bedding structure of sedimentary rocks, with layer thickness accuracy controlled within ±30 micrometers; and an isotope dating unit integrated into the geochronology laboratory, using thermal ionization mass spectrometry (TIMS) to obtain uranium-lead isotope ratios and applying a Bayesian correction algorithm (Bacon). v2.5) compresses the absolute age error to ±0.45 million years; in addition, the system is connected to a database server of extant closely related species, which contains high-precision anatomical CT scan data and motion capture trajectories of more than 1,200 species of birds and reptiles, with a sampling frequency of 240 Hz and joint angle accuracy of ±1.2 degrees.

[0034] The knowledge graph construction and storage subsystem is deployed on a distributed graph database cluster (based on Neo4j Enterprise 5.12) and uses an attribute graph model to organize data. Node types include four categories: fossil specimen nodes (including unique ID, collection coordinates, and preservation status), stratigraphic unit nodes (including lithology, sedimentary facies, and thickness), isotope sample nodes (including elemental ratios and corrected age), and extant analogous species nodes (including taxonomic information, body shape parameters, and movement pattern labels). Edge types are defined as: "excavated from" (connecting fossils and strata), "chronologically belonging to" (connecting fossils or strata and isotope samples), "anatomically homologous" (based on BLAST+sequence alignment and morphological similarity calculation, threshold > 0.78), and "motor similarity" (based on Dynamic Time Warping (DTW) distance < 0.15). This graph supports millisecond-level queries of millions of nodes and tens of millions of relationships, and provides services externally through a GraphQL interface. The spatiotemporal alignment module runs on a dedicated FPGA acceleration card (Xilinx Alveo U280) and executes a four-dimensional coordinate transformation algorithm: transforming point cloud data ( ) and microCT tomography ), isotope age ( The alignment is uniformly mapped to the WGS84-GEOID spatiotemporal framework based on the Earth reference ellipsoid, and the alignment error is less than 0.42 mm after optimization by ICP (Iterative ClosestPoint).

[0035] The AI ​​modeling and simulation subsystem serves as the core processing unit, consisting of two NVIDIA DGX A100 servers, each equipped with eight A100 80GB GPUs, 2TB of system memory, and an InfiniBand HDR 200Gb / s interconnect network. The Conditional Generative Adversarial Network (cGAN) is deployed within the TensorRT optimization engine. The generator employs an encoder-decoder architecture: the encoder contains six 3D convolutional layers, all with a kernel size of 3×3×3, a stride of 2, and channel numbers increasing sequentially from 32 to 1024, using LeakyReLU (α=0.2) as the activation function. The decoder fuses the output feature maps from each layer of the encoder via skip connections and employs transposed convolutional upsampling to ultimately output vertex coordinates (float32 format) and unit normal vectors (normalized to an L2 norm of 1). The discriminator network introduces a composite loss function. ,in Based on finite element analysis, it was calculated whether the maximum principal stress of the bone under its own weight and simulated biting force is lower than the yield strength of the cortical bone (approximately 170 MPa). Estimate the joint torque balance error by comparing the muscle lever arm with the preset muscle force; The phylogenetic tree distance matrix is ​​used to penalize the completion results that deviate too much from the morphology of closely related species. During training, the input is a fossil point cloud with a missing rate of 35% (denoised by Poisson disk sampling), and the output is a complete triangular mesh (average number of faces of about 180,000), with the geometric continuity of the completed region being C1 smooth.

[0036] The dynamic behavior simulation engine runs on PhysX 5.0 physics middleware, integrating rigid body dynamics and soft tissue deformation models. The skeletal mesh is automatically segmented into rigid body components (such as femur, tibia, and vertebral segments). Each component is assigned material parameters such as density 1.8 g / cm³, Young's modulus 18 GPa, and Poisson's ratio 0.3, and is divided into tetrahedral finite element elements with 2 mm side lengths for stress distribution calculation. The muscle attachment point inference module is based on a graph neural network (GNN). Inputs are the curvature tensor field of fossil bone surfaces and the tendon attachment heatmap of extant analogous species. Through a message passing mechanism, it predicts the origin and insertion points of muscles in extinct species with an accuracy of 93.1% (cross-validated). The joint degree-of-freedom constraint module analyzes the Gaussian curvature distribution of the joint surface point cloud, fits a spherical or ellipsoidal model, and calculates the maximum flexion / extension angle, adduction / abduction angle, and rotation angle, with an error controlled within ±4.3 degrees. The gait pattern matching module extracts 12-dimensional features, including gait frequency, support phase ratio, and center of gravity fluctuation amplitude, from a built-in database of 200 extant quadrupedal animal movements (including horses, crocodiles, kangaroos, etc.). It uses cosine similarity to retrieve the nearest neighbor template. When the similarity is ≥0.86, the template is directly used to drive the simulation; otherwise, a reinforcement learning agent is activated to autonomously explore the energy-optimal gait within the MuJoCo environment. The entire simulation runs stably at 60 frames per second, with a single frame computation time ≤16.5 milliseconds.

[0037] The multimodal interaction subsystem consists of three parts: a natural language understanding interface, a gesture recognition module, and a virtual reality visualization interface. The natural language interface is deployed on a local edge server (Intel Xeon Silver 4310 + 64GB RAM), running a Chinese paleontology terminology parsing model fine-tuned based on BERT-large. The vocabulary covers all entries in the *Terminology of Vertebrate Paleontology*, supporting 512 structured commands such as "lengthen the 12th caudal vertebra by 15%" and "increase the angle between the scapula and thoracic vertebrae to 65 degrees," achieving an F1 score of 0.94 for intent recognition. The gesture recognition module uses an Azure Kinect DK depth camera at 30fps, with a depth accuracy of ±2mm@2m. It tracks 21 key points of the hand (including fingertips, palm, and knuckles) in real time using the MediaPipe Hands model, recognizing four types of operations: zoom (changes in the distance between the hands), translation (movement of the center of the palm), sectioning (drawing a line with the index finger), and annotation (clicking with a clenched fist), achieving a recognition accuracy of 98.3%. The virtual reality interface is based on the Varjo XR-4 headset, with dual Micro-OLED displays featuring a resolution of 3840×2720 per eye, a refresh rate of 90Hz, a horizontal field of view of 110 degrees, and optical latency of less than 18 milliseconds after asynchronous time warp (ATW) compensation. It allows users to observe the restored model while circling in a 1:1 scale virtual Jurassic forest and trigger ecological rationality assessments (such as whether the vegetation height matches the feeding posture) via the controller.

[0038] The scientific verification and iterative optimization subsystem continuously monitors the reconstruction quality. The geometric overlap index, calculated using Hausdorff distance, determines the maximum deviation between the original fossil point cloud and the corresponding area in the reconstruction model, requiring ≤0.8mm (<0.001 after normalization to body length). The energy efficiency index, defined as the ratio of the estimated metabolic energy required per unit distance of movement to the mean of extant analogous species, must be between 0.85 and 1.15. Niche suitability, based on the MaxEnt model, takes paleoenvironmental reconstruction data (temperature, precipitation, vegetation type) and reconstructed body shape parameters as input, and outputs a habitat suitability probability, requiring ≥0.91. An alarm is triggered if any of these three indicators fails to meet the standards.

[0039] Based on the above system architecture, the workflow of this embodiment is as follows:

[0040] First, the data acquisition subsystem simultaneously acquires multi-source data of the target fossil: a laser confocal scanner completes surface point cloud acquisition (approximately 12 minutes), a micro-CT device reconstructs the three-dimensional structure of the surrounding rock strata (scanning time 45 minutes), and the geological laboratory submits a Bayesian-corrected isotopic dating report (error ±0.42 Ma). Simultaneously, the system automatically retrieves the five closest bird species (e.g., emu, cassowary) and three crocodiles from the extant analogy database, downloading their skeletal CT and gait data. All raw data, after preprocessing, is then sent to the knowledge graph construction module.

[0041] Subsequently, the spatiotemporal alignment algorithm is activated: the FPGA acceleration card will adjust the point cloud coordinates ( ) and microCT slices Perform vertical registration and use the deposition rate model to... Convert to relative dates Then, by combining the isotopic absolute age t_abs, a linear interpolation function t(z) is established, and finally, each point cloud vertex is assigned a four-dimensional coordinate. Meanwhile, the "anatomical homology" relationships determined based on phylogenetic analysis were written into the graph database, forming an initial knowledge graph.

[0042] Next, the AI ​​modeling subsystem loads the knowledge graph, triggering the execution of the conditional generative adversarial network. The generator takes the fossil point cloud (missing the anterior half of the head and the right forelimb) as input and outputs a complete mesh; the discriminator simultaneously calculates biomechanical losses—for example, if the added humerus is too long, causing the shoulder joint torque to exceed the range that the maximum muscle contraction force can balance, a negative gradient is fed back to force the generator to adjust. After 12 rounds of adversarial training (approximately 1.8 seconds per round), an initial 3D morphological model that satisfies anatomical consistency and biomechanical rationality is output.

[0043] The model was then imported into a dynamic behavior simulation engine. The muscle attachment point inference module predicted the origin and insertion positions of 42 major muscles based on a GNN; the joint degrees of freedom module calculated the maximum hip flexion angle to be 112°±4° and the knee extension limit to be 175°; the gait matching module found a similarity of 0.89 with the emu's walking pattern, and therefore adopted its stride frequency (1.8 Hz) and stride length (1.2 m) as initial parameters. The simulation engine ran a 10-second virtual walking test at 60fps, generating a dynamic sequence including skeletal pose, muscle activation timing, and ground reaction forces.

[0044] Meanwhile, researchers wore Varjo XR-4 headsets to enter a virtual reality environment and observe the model's walking posture in a simulated riverbank sedimentary environment. Through voice commands such as "increase tailbone swing amplitude to improve balance," the system analyzed the commands and modified the tailbone muscle activation strategy; by drawing a cross-sectional plane with a right-hand gesture, the internal structure of the pelvic girdle bones was displayed in real time. All interactions were recorded and converted into parameter adjustment commands.

[0045] Once a user confirms the effectiveness of a modification, or the lab adds a new piece of related fossil evidence (such as a newly discovered toe bone), the real-time iterative optimization mechanism is immediately activated: the incremental learning module employs a knowledge distillation strategy, using the new fossil point cloud as a soft label to fine-tune the original cGAN generator, updating only the weights of the last two convolutional kernels (accounting for 4.7% of the total parameters), ensuring that historical knowledge is not catastrophically forgotten. Simultaneously, the knowledge graph automatically inserts new nodes and updates the "excavated from" and "anatomically related" edges. The entire correction process—including graph updates, model fine-tuning, simulation reruns, and VR re-rendering—is completed within 2.8 seconds, far below the 3-second threshold.

[0046] Ultimately, the scientific validation module automatically evaluated the new model: geometric overlap 0.76mm (meeting the standard), energy efficiency index 0.93 (meeting the standard), and niche matching degree 0.94 (meeting the standard). The system determined it to be a valid restoration and stored this version (including all interaction logs, parameter snapshots, and validation indicators) in the knowledge graph, forming a traceable digital asset. The entire first complete restoration process took 7.6 seconds and supports up to 12 researchers concurrently operating different parts of the same model through their respective terminals.

[0047] This embodiment fully demonstrates the deep integration of the present invention in terms of system architecture and methodological process: the hardware platform provides computing power and interface guarantees for multi-source data fusion, AI modeling, physical simulation and multimodal interaction, while the method logic defined in step S precisely drives the collaborative work of each hardware module to form a closed-loop intelligent restoration system of "perception-modeling-simulation-interaction-evolution".

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for real-time 3D reconstruction of paleontology based on artificial intelligence and multimodal interaction, characterized in that, Includes the following steps: A unified representation framework for multi-source heterogeneous data is constructed. High-precision three-dimensional point cloud data of paleontological fossils, sedimentary rock bedding information, isotopic dating data, and anatomical and kinematic parameters of modern closely related species are collected. The multi-source heterogeneous data are mapped to a unified four-dimensional spatiotemporal coordinate system through a spatiotemporal alignment algorithm to form a structured paleontological knowledge graph. An initial three-dimensional morphological model is generated. Based on the structured paleontological knowledge graph, a conditional generative adversarial network is used to perform semantically perceptive geometric completion of the missing parts of the fossil. The generator takes the fossil point cloud as the conditional input, and the discriminator integrates biomechanical constraints and evolutionary rationality rules to output a complete three-dimensional morphological mesh with anatomical consistency. A dynamic behavior simulation engine is constructed, and the complete three-dimensional morphological mesh is imported into the physical simulation environment. Combined with the muscle attachment point inference module, joint degree of freedom constraint module and gait pattern matching module, the movement posture of ancient organisms under different terrains and speeds is simulated to generate dynamic behavior sequences that conform to biomechanical principles. To achieve multimodal human-computer collaborative interaction, a natural language understanding interface, a gesture recognition module, and a virtual reality visualization interface are deployed, allowing researchers to adjust restoration parameters through voice commands, rotate or cut models through gesture operations, and verify the ecological rationality of restoration results through immersive virtual reality environments. A real-time iterative optimization mechanism is implemented to dynamically update the structured paleontological knowledge graph based on user interaction feedback and new fossil evidence. This triggers an incremental learning process involving a conditional generative adversarial network and a dynamic behavior simulation engine. The model is then corrected and re-rendered within a preset time period, achieving closed-loop real-time restoration.

2. The method for real-time 3D reconstruction of paleontology based on artificial intelligence and multimodal interaction according to claim 1, characterized in that, The generator of the conditional generative adversarial network adopts an encoder-decoder architecture. The encoder consists of multiple layers of three-dimensional convolutional layers, and the decoder fuses multi-scale features through skip connections to output vertex coordinates and normal vectors. The discriminator introduces a biomechanical loss function that includes a skeletal strength constraint, a muscle torque balance term, and an evolutionary distance penalty term.

3. The method for real-time 3D reconstruction of paleontology based on artificial intelligence and multimodal interaction according to claim 1, characterized in that, The muscle attachment point inference module uses a graph neural network based on a database of tendon distribution in modern birds and reptiles to predict the origin and insertion points of muscles in extinct species; the joint degree of freedom constraint module calculates the maximum flexion and extension angles based on the curvature of fossil articular surfaces; and the gait pattern matching module retrieves the most similar gait template from a extant tetrapod motion database.

4. The method for real-time 3D reconstruction of paleontology based on artificial intelligence and multimodal interaction according to claim 1, characterized in that, The natural language understanding interface supports the parsing of Chinese professional terms and the recognition of multiple paleontological academic commands; the gesture recognition module captures key points of the hand based on a depth camera and recognizes four basic operations: zoom, pan, slice, and annotation; the virtual reality visualization interface meets the refresh rate, field of view, and latency performance standards for immersive interaction.

5. The method for real-time 3D reconstruction of paleontology based on artificial intelligence and multimodal interaction according to claim 1, characterized in that, The incremental learning process employs a knowledge distillation mechanism, injecting new evidence as soft labels into the original model. This retains historical knowledge while absorbing new information, and the amount of model parameter updates is controlled within a predetermined proportion of the total number of parameters.

6. The method for real-time 3D reconstruction of paleontology based on artificial intelligence and multimodal interaction according to claim 1, characterized in that, The structured paleontological knowledge graph is stored using an attribute graph model. Node types include fossil specimens, stratigraphic units, isotope samples, and extant analogous species. Edge types include "excavated from", "belongs to", "anatomically similar", and "motorally similar".

7. The method for real-time 3D reconstruction of paleontology based on artificial intelligence and multimodal interaction according to claim 1, characterized in that, The dynamic behavior simulation engine integrates rigid body dynamics and soft tissue deformation models. The skeletal part adopts finite element analysis, while the soft tissue part adopts a mass-spring system, with the spring constant dynamically adjusted according to the muscle type.

8. The method for real-time 3D reconstruction of paleontology based on artificial intelligence and multimodal interaction according to claim 1, characterized in that, It also includes scientific verification steps, automatically calculating the geometric overlap between the restoration model and the original fossil data, the energy efficiency index of the movement trajectory, and the ecological niche matching degree. All three indicators must be greater than or equal to the preset validity threshold to be considered a valid restoration.

9. A real-time 3D paleontological reconstruction system based on artificial intelligence and multimodal interaction, characterized in that: The system comprises a data acquisition subsystem, a knowledge graph construction and storage subsystem, an AI modeling and simulation subsystem, a multimodal interaction subsystem, and a scientific verification and iterative optimization subsystem. The data acquisition subsystem acquires high-precision 3D point cloud data of paleontological fossils, sedimentary rock bedding information, isotopic dating data, and anatomical and kinematic parameters of modern closely related species. The knowledge graph construction and storage subsystem maps multi-source heterogeneous data to a unified four-dimensional spatiotemporal coordinate system using a spatiotemporal alignment algorithm to form a structured paleontological knowledge graph. The AI ​​modeling and simulation subsystem generates an initial 3D morphological model based on the structured paleontological knowledge graph using a conditional generative adversarial network (GAN) and constructs a dynamic behavior simulation engine to generate dynamic behavior sequences. The multimodal interaction subsystem deploys a natural language understanding interface, a gesture recognition module, and a virtual reality visualization interface to achieve multimodal human-computer collaborative interaction. The scientific verification and iterative optimization subsystem dynamically updates the structured paleontological knowledge graph based on user interaction feedback and new fossil evidence, triggering an incremental learning process to complete model correction and re-rendering.

10. The real-time three-dimensional paleontological reconstruction system based on artificial intelligence and multimodal interaction according to claim 9, characterized in that, The AI ​​modeling and simulation subsystem is deployed on a GPU cluster, and the multimodal interaction subsystem is connected to the cloud inference server via a high-speed wireless network. The time taken for a single complete restoration process is less than the preset time limit, and it supports concurrent collaboration among multiple researchers.

Images