A Generative Design Approach for New Material Formulations in Additive Manufacturing

Abstract

This invention relates to the field of intelligent manufacturing technology, specifically to a generative design method for novel material formulations for additive manufacturing. The method includes: using a generative adversarial network model to learn the complex mapping relationships between the components, process parameters, and performance indicators of known biomaterials, and proactively generating novel material formulations or gradient formulations that meet target performance; rapidly verifying and evaluating the generated formulations using a meta-learning model combined with a small amount of key experimental data, and feeding back the evaluation results to optimize the generative model; integrating physics-based simulations for manufacturability verification, and combining multi-objective optimization and user interaction to determine the optimal formulation; finally, converting the formulation into equipment instructions, which can be adaptively adjusted in conjunction with real-time process monitoring. This invention can significantly accelerate the research and development process of novel materials for additive manufacturing, especially biomedical composite materials, reduce research and development costs, and improve the performance consistency and reliability of the final product.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and specifically to a novel material formulation generative design method for additive manufacturing. Background Technology

[0002] Additive manufacturing, particularly for biomedical materials, has provided a revolutionary means for the fabrication of personalized, complex medical devices. Its core lies in materials, whose properties directly determine the biological function and mechanical reliability of the final product. However, the formulation development of new materials for additive manufacturing, especially composite materials or functionally graded materials, still faces significant challenges. Traditional new material development relies heavily on researchers' experience and extensive trial-and-error experiments, screening for solutions that meet target performance from a large number of candidate components and process parameter combinations. This process is time-consuming, costly, and highly uncertain. In recent years, computational simulation and machine learning methods have been introduced into materials research, for example, by establishing performance prediction models to accelerate the screening process. However, these methods mostly belong to the "forward prediction" mode, i.e., predicting performance given a formulation. Essentially, they still involve a limited exploration of the known design space, making it difficult to proactively generate innovative formulations that transcend existing knowledge. Furthermore, there is a highly nonlinear and complex relationship between material properties and composition and processes, which a single model cannot accurately capture. In additive manufacturing scenarios, the manufacturability of formulations is crucial. However, existing design methods are often disconnected from manufacturing processes. The designed formulations may be theoretically feasible, but they cannot be stably formed in actual printing. Therefore, there is an urgent need in this field for a new paradigm that can deeply integrate artificial intelligence, proactive design, manufacturability constraints, and efficient experimental verification to systematically solve the integrated intelligent design problem from performance targets to printable formulations, and break through the bottleneck of new material research and development.

[0003] Therefore, existing technologies still need further development. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide a new material formulation generative design method for additive manufacturing, so as to solve the problems existing in the prior art.

[0005] To achieve the above-mentioned technical objectives, this invention provides a novel material formulation generative design method for additive manufacturing, comprising:

[0006] S1. Utilize generative adversarial network models to learn the mapping relationship between the composition, process parameters, and performance indicators of known biomaterials;

[0007] S2. Generate new material formulations or gradient formulations for additive manufacturing processes based on the mapping relationship;

[0008] S3. The generated formula is validated and fine-tuned using a small amount of experimental data through a meta-learning model.

[0009] Specifically, the method of using a generative adversarial network model to learn the mapping relationship between the composition, process parameters, and performance indicators of known biomaterials includes:

[0010] A training dataset was constructed using known biomaterial composition data, process parameter data, and performance index data. The process parameter data included at least one of the printing temperature, printing speed, layer thickness, and curing parameters in the additive manufacturing process. A conditional generative adversarial network architecture was adopted, in which the target performance index was simultaneously input into the generator and discriminator as a conditional vector. This enabled the generator to output candidate formulations that met the target performance index, while the discriminator evaluated the consistency between the generated formulations and the real formulations under given conditions. The network weights of the generator and discriminator were iteratively optimized through an adversarial training process until the generator could stably output formulations that met the performance constraints. This training process integrated a gradient penalty loss function to improve training stability.

[0011] Specifically, the generator adopts an encoder-decoder structure. The encoder maps the component data and process parameter data into a low-dimensional latent space representation, and the decoder reconstructs or generates new formulations from this latent space representation. A multi-head self-attention mechanism is embedded in the encoder-decoder to capture the cross-modal dependencies between components and process parameters and to focus on the impact of key features on performance indicators. During training, spectral normalization technology is introduced to constrain the weight matrix of the discriminator, and a mini-batch discrimination mechanism is combined to increase the diversity of generated formulations, effectively alleviate the mode collapse problem, and ensure the continuity and explorability of generated formulations in the latent space.

[0012] Specifically, the decoder outputs a formula including the mixing ratio of various bioactive materials, gradient distribution function, and recommended additive manufacturing process parameter adjustment values. The formula is represented in structured vector or matrix form and can be directly converted into an input format that the additive manufacturing equipment can parse. After generating the formula, a physics-based simulation model is integrated to verify its manufacturability. The simulation model simulates the flow, solidification, and interlayer bonding behavior of materials during the printing process, predicts possible porosity, cracks, or deformation defects, and feeds the simulation results back to the generator to adjust the formula generation strategy, ensuring the theoretical printability of the formula.

[0013] Specifically, the step of using a meta-learning model to validate and fine-tune the generated formula with a small amount of experimental data includes:

[0014] A model-independent meta-learning framework is adopted to perform meta-training on multiple heterogeneous material formulation verification tasks to learn an initial model parameter that adapts quickly. For each newly generated formulation, a small number of key experiments are designed and executed to obtain its performance data, forming a support set. The meta-learning model is quickly adapted using the support set data through a small number of gradient descent steps to obtain a task-specific model. This model is used to predict the unmeasured performance of the formulation or evaluate the deviation from the target performance. The loss gradient is calculated based on the deviation signal to fine-tune the parameters of the generative adversarial network model.

[0015] Specifically, the meta-training process of the model-independent meta-learning framework includes:

[0016] The model samples support and query sets across multiple tasks. On the support set, it computes task-specific losses and updates model parameters. On the query set, it evaluates the meta-loss and computes meta-gradients with respect to initial parameters. The meta-optimizer updates the initial parameters to enable the model to adapt quickly to new tasks. During the validation phase, it actively selects experimental points using a Bayesian optimization strategy to maximize expected improvement or minimize uncertainty, thereby efficiently acquiring the most informative experimental data for model adaptation. The meta-learning model also integrates an uncertainty quantification module to provide confidence estimates for performance prediction, guiding early cessation or further exploration in the validation process.

[0017] Specifically, the generation of new material formulations or gradient formulations for additive manufacturing processes also includes a multi-objective optimization process:

[0018] Multiple competitive performance objectives are defined, including biocompatibility, degradation rate, mechanical strength, printing accuracy, and material cost. The potential space of the generative adversarial network is used as the design space, and an evolutionary algorithm based on non-dominated ranking or a Pareto front optimization algorithm is used to explore it, generating a series of non-dominated solutions to form a Pareto optimal formulation set. An interactive interface is provided to allow users to specify preferences or weights, select one or more candidate formulations from the Pareto front according to preferences, and use this selection information to further constrain the conditional input of the generative adversarial network to generate formulations that better meet the user's equilibrium point.

[0019] Specifically, this also includes integrating the generated formula with the additive manufacturing equipment control system:

[0020] A dedicated conversion module converts formula data and its associated process parameters into control instructions executable by the equipment, including G-code, slice files, or direct drive signals. During the formula generation stage, an integrated equipment hardware constraint model is used, which includes nozzle diameter, heating temperature range, platform motion accuracy, and material delivery system limitations, ensuring that the generated formula is executable within the physical limits of the equipment. A closed-loop control process is established, and virtual simulation verification is performed during the printing preparation stage. The formula or process parameters are then fine-tuned based on the simulation results.

[0021] Specifically, the closed-loop control process also includes real-time online adjustment:

[0022] In the additive manufacturing printing process, sensors integrated into the equipment are used to monitor key process variables, including extrusion pressure, temperature distribution, layer thickness deviation, and degree of curing. The real-time data stream is input into a pre-trained meta-learning model, which dynamically evaluates the deviation between the current printing state and the expected state and generates adjustment signals. The adjustment signals are fed back to the online fine-tuning module of the generative adversarial network model to adjust the material formulation ratio or process parameter settings of subsequent printing layers in real time, so as to achieve adaptive printing and performance compensation and ensure the consistency of the final part performance.

[0023] Specifically, this also includes establishing a continuous iterative optimization loop:

[0024] After each experimental verification or actual printing of the generated formula, the obtained performance and process data are collected and labeled, and expanded into the historical training database. The generative adversarial network model and meta-learning model are retrained periodically using the expanded data, and their parameters are updated to absorb new knowledge. This cycle enables the system to adapt to the introduction of new material systems, changes in process parameters, or changes in performance targets, gradually improving the first success rate of formula generation and expanding its designable material space, forming a self-evolving generative design system for new material formulas.

[0025] Beneficial effects:

[0026] Compared with existing technologies, the novel material formulation generative design method for additive manufacturing provided by this invention has the following significant advantages:

[0027] First, this invention, by integrating generative adversarial networks (GANs) and meta-learning, constructs an intelligent closed loop of "active design-efficient verification," fundamentally changing the paradigm of new material development. GANs can deeply mine the complex nonlinear mapping relationships between "composition-process-performance" from historical data and proactively generate novel formulations that satisfy multi-objective constraints. This breaks through the passive screening limitations of traditional methods confined to existing data, enabling creative exploration of a vast unknown design space. The introduction of the meta-learning model allows the system to leverage the generalization ability gained from historical tasks—learning "how to learn"—to quickly construct and verify accurate local performance prediction models with only a minimal number of key experiments for newly generated formulations. This reduces the massive number of experiments required for verification to single digits, significantly lowering the trial-and-error costs and time cycle of R&D, and achieving intelligent and efficient material development.

[0028] Secondly, this invention ensures the feasibility and practicality of the design results through multi-dimensional and multi-stage constraints and optimizations. During the generation stage, a physics-based simulation model is integrated for manufacturability verification, predicting and mitigating potential flow defects and warping during printing, achieving "manufacturability-driven design" and avoiding invalid solutions that cannot be printed. Furthermore, a multi-objective optimization algorithm explores the trade-offs between performance parameters and presents the Pareto front visually, supporting interactive decision-making by users based on actual application scenarios, making the design solution more aligned with actual engineering needs. A hardware constraint model is introduced to ensure that the generated formula parameters are strictly within the physical limits of the target equipment. These progressive measures guarantee a seamless connection and high reliability between the theoretical formula generated by the intelligent algorithm and the final stable manufacturing solution.

[0029] Finally, this invention achieves self-adaptation in the manufacturing process and autonomous growth of system knowledge by constructing a self-evolving system that incorporates real-time online adjustment and continuous iterative learning. The real-time online adjustment module utilizes sensor data and pre-trained models to dynamically sense state deviations during printing and fine-tune process parameters in real time, compensating for uncertainties in the manufacturing process and significantly improving the consistency and reliability of the final part's performance. Meanwhile, the continuous iterative optimization loop automatically transforms data generated from each experiment and manufacturing process into system knowledge, periodically updating the core model. This allows the system to continuously adapt to new materials, processes, and equipment, gradually expanding its designable scope and improving the first-time success rate. This makes this invention not only a design tool but also an intelligent design and manufacturing ecosystem with self-improvement and growth capabilities, providing a sustainable solution for the efficient and reliable development of new additive manufacturing materials. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating the generative design method for new material formulations for additive manufacturing provided in a specific embodiment of the present invention. Detailed Implementation

[0031] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments in this application, other similar embodiments obtained by those skilled in the art without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0032] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0033] Please see Figure 1 This invention provides a novel material formulation generative design method for additive manufacturing, comprising:

[0034] S1. Use a generative adversarial network model to learn the mapping relationship between the composition, process parameters and performance indicators of known biomaterials.

[0035] It should be further explained that this invention provides a complete framework for a novel material formulation generative design method for additive manufacturing. The core of this method lies in constructing an intelligent material design closed loop. First, the system needs to establish a database containing historical data, which includes detailed records of known biomaterials. Each record includes at least the material composition, preparation process parameters, and the final measured performance indicators. The material composition may include, but is not limited to, the mass percentage or molar ratio of one or more biomaterials such as polylactic acid, polycaprolactone, β-tricalcium phosphate, hydroxyapatite, collagen, and silk fibroin. Process parameters must be closely related to the additive manufacturing process. For example, the printhead temperature in fused deposition modeling (FDM) typically ranges from 150°C to 250°C, chosen because it covers the melting temperature of most biodegradable polymers; the printing speed typically ranges from 10 mm / s to 100 mm / s, balancing printing efficiency and forming accuracy; the layer thickness typically ranges from 0.1 mm to 0.3 mm, a common setting to ensure interlayer bonding strength and surface accuracy; and the exposure time and light intensity in photopolymerization. Performance indicators are set according to application goals. For example, the compressive modulus of scaffold materials used for bone repair may be between 50 MPa and 500 MPa to match the mechanical properties of cancellous bone; the degradation cycle target may be 3 to 12 months; and biological properties such as cell viability.

[0036] S2. Generate new material formulations or gradient formulations for additive manufacturing processes based on the mapping relationship.

[0037] It should be further explained that the specific design of the above scheme includes:

[0038] (1) In the data preparation stage, all data need to be normalized or standardized preprocessed to eliminate the influence of units and accelerate model convergence. For example, all numerical features can be scaled to the [0,1] interval. Then, the processed dataset is divided into training, validation and test sets in a ratio of approximately 7:1.5:1.5. This ratio is a common division in the field of machine learning, which aims to ensure that the model has sufficient data to learn from, while having independent datasets for verifying generalization ability and final evaluation.

[0039] (2) One of the core models of the system is the Generative Adversarial Network (GAN). The GAN consists of a generator G and a discriminator D. During training, the generator G learns to generate a realistic recipe data distribution P_G that satisfies condition c from a random noise vector z and a target condition vector c (representing the desired performance metric, such as "compression modulus > 200 MPa"). The discriminator D learns to distinguish between real recipe data x from the real data distribution P_data and fake data G(z|c) generated by the generator G, and determines whether it satisfies condition c. The training objective is to minimize the difference between the generated data and the real data distribution, which is a min-max game process. Its objective function V(D,G) can be formalized as:

[0040]

[0041] in, Representing formula data from a real data distribution. It is the distribution of real formula data. It comes from the prior distribution A random noise vector (usually a standard normal distribution). It is a conditional vector representing the target performance. This indicates that the discriminator judges the true data. Under conditions The following is the probability of it being true. Indicates that the generator uses noise and conditions To input the generated recipe data, This represents the probability that the discriminator determines the generated data to be true. It represents the mathematical expectation.

[0042] (3) To improve the stability of generative adversarial network training, a gradient penalty term is introduced into the loss function of the discriminator D, forming the Wasserstein GAN with Gradient Penalty architecture. Gradient penalty term The form is ,in In real data points and generate data points Randomly sampled points on the connecting line, yes The distribution It is a discriminator pair gradient, It is the L2 norm. This is the penalty coefficient, usually set to 10. Increasing this penalty term can effectively prevent gradient explosion or vanishing in the discriminator, making the training process smoother.

[0043] (4) After receiving a new target performance condition vector c_new, the fully trained generator G can sample from the noise distribution to generate a series of new material formulation candidates that are “reasonable” in terms of data distribution and whose target is c_new. These formulation candidates not only include the proportion of material components, but also the recommended process parameters (such as “at a printing temperature of 200°C”), and are output in the form of structured vectors for easy subsequent processing.

[0044] S3. The generated formula is validated and fine-tuned using a small amount of experimental data through a meta-learning model.

[0045] It should be further noted that recipes generated solely by generative adversarial networks are still based on statistical extrapolation from historical data, and their actual performance needs to be verified experimentally. To reduce experimental costs, a meta-learning model is integrated into the system. This meta-learning model is trained on a large number of different material formulation-related tasks to "learn how to learn," thereby obtaining well-defined initial parameters. When faced with a new generative recipe verification task, only a small number of experiments (e.g., 3-5) are needed to obtain performance data for the recipe or its minor variants. Based on this limited data, the meta-learning model can be quickly fine-tuned (typically 5-10 steps of gradient descent) to obtain a predictive model adapted to the specific recipe verification task. This model can predict the performance of the recipe under other untested conditions or assess its gap with the target performance.

[0046] Furthermore, this invention uses the performance gap evaluated by the meta-learning model as a feedback signal, which is backpropagated to the generator G of the generative adversarial network to fine-tune its parameters, thereby guiding the next round of generation of a recipe that is closer to the target and more likely to be experimentally validated. This forms a closed loop of "generation-validation-feedback-optimization," iterating continuously until a final recipe that meets all performance objectives and passes validation is obtained.

[0047] Understandably, this invention combines the powerful generative capabilities of generative adversarial networks with the rapid adaptation capabilities of meta-learning to small sample sizes, providing an automated intelligent solution for the design of new material formulations for additive manufacturing. This method can uncover potential patterns from complex "composition-process-performance" relationships, proactively explore new formulation spaces that are difficult to reach with traditional trial-and-error methods, and significantly shorten the cycle from design to verification of new materials by utilizing an efficient, low-volume experimental verification strategy, thereby reducing R&D costs. It is a powerful tool for accelerating the development of advanced biomedical materials.

[0048] Specifically, the method of using a generative adversarial network model to learn the mapping relationship between the composition, process parameters, and performance indicators of known biomaterials includes:

[0049] A training dataset was constructed using known biomaterial composition data, process parameter data, and performance index data. The process parameter data included at least one of the printing temperature, printing speed, layer thickness, and curing parameters in the additive manufacturing process. A conditional generative adversarial network architecture was adopted, in which the target performance index was simultaneously input into the generator and discriminator as a conditional vector. This enabled the generator to output candidate formulations that met the target performance index, while the discriminator evaluated the consistency between the generated formulations and the real formulations under given conditions. The network weights of the generator and discriminator were iteratively optimized through an adversarial training process until the generator could stably output formulations that met the performance constraints. This training process integrated a gradient penalty loss function to improve training stability.

[0050] It should be further noted that constructing the training dataset is a prerequisite for the model's success. Data should originate from public databases, published literature, and internal experimental reports from collaborating laboratories. To ensure data quality, rigorous cleaning is required, including handling missing values ​​and removing obvious outliers. For composition data, it must be standardized to mass percentage or molar percentage, with a total of 100%. Process parameters must be associated with the specific additive manufacturing equipment type. For example, for fused deposition modeling, the printing temperature record could be the nozzle temperature, the printing speed the extruder head's movement speed, and the layer thickness the height per layer set in the slicing software. A typical data entry example is as follows: Composition [Polylactic acid (PLA): 85%, β-tricalcium phosphate (β-TCP): 15%]; Process [Printing temperature: 210°C, Printing speed: 30 mm / s, Layer thickness: 0.2 mm]; Properties [Compressive strength: 45 MPa, Degradation rate (8-week mass loss): 25%, Cell proliferation rate (relative to control): 120%].

[0051] Furthermore, the specific design is as follows:

[0052] (1) The specific implementation scheme of the conditional generative adversarial network is as follows: The input of the generator G is a concatenated noise vector z and a conditional vector c. The conditional vector c is the encoding of the target performance. For example, if the target is "compression modulus greater than 200MPa and degradation period of 6-9 months", then c can be a multi-dimensional vector, such as [200, 0.6, 0.9] (the first element represents the lower limit of the modulus, and the last two represent the degradation period range). The generator G is usually composed of multiple fully connected layers or transposed convolutional layers, and the final output is a recipe vector with the same dimension as the training data. Batch normalization layers and ReLU activation functions can be used in the generator G to facilitate training.

[0053] (2) The input to the discriminator D is the concatenation of the formula data x (or the generated data G(z|c)) and the condition vector c. The discriminator D consists of multiple fully connected layers and ultimately outputs a scalar through a sigmoid activation function, representing the probability that the input data is real data and holds true under given conditions. Batch normalization is usually avoided in the discriminator D; instead, layer normalization or spectral normalization is used. During training, the discriminator D and the generator G are updated alternately. In each step, the discriminator D is updated multiple times (e.g., 5 times), and then the generator G is updated once. This training strategy helps maintain the competitiveness of the discriminator and prevents the generator from quickly "cheating" a weaker discriminator.

[0054] (3) The training process of Wasserstein GAN with integrated gradient penalty is more stable. Its implementation steps include: First, sampling a real batch of data x and corresponding condition c from the training set, and sampling noise z from the noise distribution. Then, obtaining fake data G(z|c) through a generator. Next, obtaining sampling points through random linear interpolation between the real data x and the fake data G(z|c). Calculate the gradient of the discriminator output at that point, and calculate the gradient penalty term. Total loss of the discriminator for: ,in Expressing expectations, This is the penalty coefficient, set to 10 as mentioned before. The loss of the generator G. for: These loss functions are iteratively optimized through backpropagation and an optimizer (such as Adam, with a learning rate of 0.0002, an exponential decay rate β1 for the first moment estimate of 0.5, and an exponential decay rate β2 for the second moment estimate of 0.9).

[0055] Understandably, by constructing a high-quality, multi-dimensionally correlated dataset and training it with a conditional generative adversarial network integrating gradient penalties, this method can accurately capture the complex nonlinear relationships in the field of biomaterials. Using target performance as a conditional input makes the generation process highly targeted, enabling direct formulation design based on user-defined mechanical, degradation, or bioactivity goals. This significantly improves the efficiency and purposefulness of the design process, avoiding blind searches.

[0056] Specifically, the generator adopts an encoder-decoder structure. The encoder maps the component data and process parameter data into a low-dimensional latent space representation, and the decoder reconstructs or generates new formulations from this latent space representation. A multi-head self-attention mechanism is embedded in the encoder-decoder to capture the cross-modal dependencies between components and process parameters and to focus on the impact of key features on performance indicators. During training, spectral normalization technology is introduced to constrain the weight matrix of the discriminator, and a mini-batch discrimination mechanism is combined to increase the diversity of generated formulations, effectively alleviate the mode collapse problem, and ensure the continuity and explorability of generated formulations in the latent space.

[0057] It should be further noted that the generator employs an encoder-decoder structure, particularly variants based on Transformer or integrated self-attention mechanisms, which can better handle the complex interactions between multiple components in material formulations. Specific designs include:

[0058] (1) The encoder part consists of multiple identical layers stacked together, each layer containing a multi-head self-attention sublayer and a feedforward neural network sublayer. The input data (recipe vector) is first converted into a vector sequence through the embedding layer. In the self-attention mechanism, the query vector for each input position (corresponding to a feature in the recipe, such as the content of PLA) is... Key vector Sum value vector Through linear transformation, we obtain: , , ,in It is a matrix representation of the input sequence. , , It is a learnable weight matrix. Attention weights are... Calculation, where This is the dimension of the key vector, used to scale the dot product and prevent gradient vanishing. Multi-head attention allows the model to simultaneously focus on information from different representation subspaces; its output is the concatenation of multiple head outputs followed by a linear transformation. The encoder's final output is a low-dimensional dense representation of the input recipe in the latent space. .

[0059] (2) The decoder part also consists of multiple stacked layers. In addition to the self-attention sublayer and the feedforward sublayer, each layer also contains an additional cross-attention sublayer, which is used to focus on the encoder output. The decoder uses a conditional vector c and / or random noise. Using the initial input, a self-attention layer is used to establish dependencies within the target sequence, and then a cross-attention layer is used to extract the latent representation from the encoder. Relevant information is extracted and a new recipe sequence is generated through a feedforward layer and an output layer. This structure enables the generator not only to learn the statistical distribution of the recipe but also to understand the causal relationships between features within the recipe. For example, increasing the β-TCP content usually requires adjusting the printing temperature to prevent clogging.

[0060] (3) To prevent pattern collapse (where the generator only generates a limited number of recipes), a mini-batch discrimination mechanism is introduced into the discriminator. Specifically, the discriminator not only judges individual samples but also calculates a certain statistical feature of all samples within a mini-batch (such as calculating the L1 distance between samples using a learnable tensor). This feature is then concatenated with the features of the sample itself before being input into subsequent layers of the discriminator network for final judgment. This allows the discriminator to perceive the diversity of the entire mini-batch data, thereby forcing the generator to generate diverse samples.

[0061] (4) For each layer of the discriminator D, the weight matrix Applying spectral normalization, i.e. ,in yes The spectral norm (maximum singular value) can be approximated using the power iteration method. Spectral normalization strictly restricts the discriminator function to a 1-Lipschitz continuous space, which is a requirement of Wasserstein distance theory. This fundamentally improves the stability of generative adversarial network training and complements gradient penalty techniques.

[0062] Understandably, the encoder-decoder structure combined with a self-attention mechanism endows the model with powerful feature extraction and relation modeling capabilities, enabling a deep understanding of complex constraints such as "high hydroxyapatite content requires lower printing speeds to ensure uniform extrusion." The combination of spectral normalization and mini-batch discrimination, while ensuring training stability, greatly promotes the diversity of generated formulations, avoids design results getting trapped in local optima, and ensures the breadth and depth of design space exploration, which is crucial for discovering innovative material combinations.

[0063] Specifically, the decoder outputs a formula including the mixing ratio of various bioactive materials, gradient distribution function, and recommended additive manufacturing process parameter adjustment values. The formula is represented in structured vector or matrix form and can be directly converted into an input format that the additive manufacturing equipment can parse. After generating the formula, a physics-based simulation model is integrated to verify its manufacturability. The simulation model simulates the flow, solidification, and interlayer bonding behavior of materials during the printing process, predicts possible porosity, cracks, or deformation defects, and feeds the simulation results back to the generator to adjust the formula generation strategy, ensuring the theoretical printability of the formula.

[0064] It needs to be further clarified that the decoder output needs to be precisely parsed into an executable design scheme. The output can be a high-dimensional vector, with dimensions corresponding to all possible components and process parameters. For example, for a system supporting three basic materials (A, B, C) and four process parameters, the output vector might be [A%, B%, C%, temperature, velocity, layer thickness, fill rate], where the percentages are processed by softmax to ensure a sum of 100%. For gradient formulations, the output can be a function parameter with respect to spatial location (such as Z-axis height), such as "along the build direction, the volume fraction of material A linearly changes from 100% at the bottom to 0% at the top, while material B linearly changes from 0% to 100%". This function can be recognized by slicing software and used to generate the corresponding gradient G-code. Specific designs include:

[0065] (1) Physics-based simulation is a key bridge connecting virtual design and real manufacturing. For material extrusion molding, computational fluid dynamics simulation of non-isothermal and non-Newtonian fluids can be integrated. The inputs to the simulation model include material rheological parameters in the generated formula (such as viscosity-shear rate curves determined by component ratios and thermal conductivity) and process parameters (printing temperature, speed, and layer thickness). By solving the mass, momentum, and energy conservation equations, the simulation can predict the flow pressure of the melt in the nozzle, the morphology of the extruded filament, the cooling and shrinkage process of the filament after deposition, and the thermal bonding between layers.

[0066] (2) Simulation can identify potential manufacturability issues. For example, if simulation shows that the extrusion pressure exceeds the nozzle's tolerance limit, or predicts insufficient interlayer bonding strength due to excessively rapid cooling of the filaments after deposition (which can be assessed by calculating the degree of healing at the contact surfaces of adjacent filaments), or predicts significant warping due to uneven shrinkage, then the formulation is deemed to have manufacturability risks. These risks are quantified into one or more "manufacturability scores". The lower the value, the more serious the problem.

[0067] (3) The simulation results are fed back to the generator of the generative adversarial network. This can be achieved in several ways. One way is to use the manufacturability score... As an additional condition or constraint, it is added to the input condition vector c of the generator G, so that while pursuing the target performance, the generator must also meet the manufacturability threshold (e.g., Another approach is to transform the simulation-predicted defect type (such as "high warpage risk") into a penalty term for certain dimensions of the recipe vector, which is then added to the loss function of the generator G. For example, if the simulation predicts that high warpage risk is related to excessively high printing temperatures, then the loss function is used to penalize excessively high temperature values ​​in the generated recipes, guiding the generator to generate recipes with more reasonable printing temperatures in subsequent iterations.

[0068] Understandably, introducing physics-based simulation verification at the generation stage enables "manufacturability-driven design." This avoids generating many theoretically superior "paper recipes" that cannot be stably manufactured using existing additive manufacturing equipment, thus narrowing the design space to a practically achievable range. This feedforward verification mechanism significantly improves design efficiency, reduces repeated modifications due to manufacturing failures, and ensures that the generated recipes not only meet performance standards but also possess good process feasibility. It is a core step in achieving integrated "design-manufacturing."

[0069] Specifically, the step of using a meta-learning model to validate and fine-tune the generated formulations with a small amount of experimental data includes: employing a model-independent meta-learning framework to perform meta-training on multiple heterogeneous material formulation validation tasks, learning a rapidly adaptable initial model parameter; for each newly generated formulation, designing and executing a small number of key experiments to obtain its performance data, forming a support set; using the support set data to rapidly adapt the meta-learning model through a small number of gradient descent steps to obtain a task-specific model, which is used to predict the unmeasured performance of the formulation or assess the deviation from the target performance, and calculating the loss gradient based on the deviation signal to fine-tune the parameters of the generative adversarial network model.

[0070] It should be further explained that meta-learning models aim to solve the problem of few-shot learning. The core idea of ​​model-independent meta-learning methods is to find a set of initial parameters for the model. This makes it possible for tasks to be distributed Any new task in the middle sampling The model only needs a small amount of support set data provided by the task. By performing a few steps (e.g., 1-5 steps) of gradient descent updates, we can obtain new parameters that perform well on this task. The specific design is as follows:

[0071] (1) The meta-training process requires the construction of a large number of meta-tasks. Each meta-task Simulate a validation process for a new formula. (Task) Includes a support set and a query set In a materials verification scenario, a task This can be used to validate different formulations within a basic material system (such as PLA / HA composites). The support set contains composition, processing, and partial performance data (small sample) for k different formulations (e.g., k=5) under this material system. The query set contains more comprehensive performance data for these k formulations or data for other related formulations, used to evaluate the predictive ability of the adapted model. The goal of meta-training is to minimize the sum of the model's losses on its respective query sets across all tasks after adaptation. Its meta-objective function is:

[0072]

[0073] in, The parameter is The base model (such as a performance prediction neural network). It is a model In the mission Loss on the support set It is the learning rate of the inner loop (i.e., the learning rate adapted within the task). These are the adapted, task-specific parameters. This represents the loss of the adapted model on the query set. The initial parameters are obtained by optimizing this meta-objective. They possess the ability to quickly adapt to new tasks.

[0074] (2) After meta-training is completed, a new formula generated by a generative adversarial network is encountered. The verification process is as follows: First, design a small number (e.g., 3) of key experiments. These experimental points can be the formulation. The study itself, and variants derived from it by making minor perturbations to key components or process parameters (e.g., ±5% component adjustments). These three experiments were performed to obtain preliminary performance data (such as compressive strength and degradation rate), forming the support set. Then, the meta-learning model is trained using the support set data. Perform rapid adaptation. The adaptation process involves executing a few steps (e.g., 3 steps) of gradient descent: The adapted model That is, for the formula Specific prediction models.

[0075] (3) Task-specific model It can be used for two things:

[0076] First, predicting the formula Performance under other untested conditions (such as mechanical behavior at different loading rates);

[0077] Second, and more importantly, assessment The deviation between the predicted performance and the target performance This deviation It can serve as a loss signal Through calculation For the parameters of the generator G in a generative adversarial network gradient And use this gradient to update the generator: ,in This is the feedback learning rate. In this way, the results of the experimental verification can be directly used to guide the generation of the next round of better formulations.

[0078] Understandably, meta-learning models transform traditional performance prediction models, which require extensive retraining with large amounts of data, into fast, plug-and-play adapters. Leveraging general "materials knowledge" learned from historical tasks, they rapidly build locally accurate prediction models for new formulations using experimental data from a very small number of new formulations. This is equivalent to gaining preliminary verification and in-depth evaluation capabilities for generated formulations at extremely low experimental costs, and seamlessly feeding the evaluation results back to the design phase, achieving a tight closed loop of "experimental verification - design optimization." This is a key technology for significantly reducing the number of experimental iterations in R&D.

[0079] Specifically, the meta-training process of the model-independent meta-learning framework includes:

[0080] The model samples support and query sets across multiple tasks. On the support set, it computes task-specific losses and updates model parameters. On the query set, it evaluates the meta-loss and computes meta-gradients with respect to initial parameters. The meta-optimizer updates the initial parameters to enable the model to adapt quickly to new tasks. During the validation phase, it actively selects experimental points using a Bayesian optimization strategy to maximize expected improvement or minimize uncertainty, thereby efficiently acquiring the most informative experimental data for model adaptation. The meta-learning model also integrates an uncertainty quantification module to provide confidence estimates for performance prediction, guiding early cessation or further exploration in the validation process.

[0081] It should be further explained that the training process of MAML is a two-layer optimization problem. The specific algorithm steps are as follows:

[0082] 1. Initialize model parameters .

[0083] 2. For iteration count = 1 to N:

[0084] a. Randomly sample a batch of meta-tasks .

[0085] b. For each task in :

[0086] c. From the task Medium sampling support set and query set .

[0087] ii. In the support set Calculate loss And perform one or more steps of gradient descent to obtain the adapted parameters: Here This is the internal learning rate, which is usually small (e.g., 0.01).

[0088] iii. In the query set Calculate loss This loss is used for meta updates.

[0089] iv. Aggregate the loss of all tasks on the query set: .

[0090] v. to calculate Regarding initial parameters Meta gradient: The second derivative needs to be calculated here, which can be achieved through automatic differentiation.

[0091] e. Use a meta-optimizer (such as Adam, with its meta-learning rate) (Typically 0.001) Update initial parameters: .

[0092] Furthermore, the specific design of the above scheme includes:

[0093] (1) In the validation phase, in order to obtain the most informative data with the fewest number of experiments to adapt to the meta-learning model, Bayesian optimization was used to actively select experimental points. The new formula... The design space (such as a small perturbation space around its composition and process parameters) is used as the optimization domain. First, an initial version of the meta-learning model is used. As a surrogate model, a Gaussian process is added to it to model the uncertainty of the prediction. Then, a sampling function (such as the expected improvement in efficiency, EI) is selected to determine the next experimental point. The formula for calculating the expected improvement in efficiency, EI, is as follows: ,in This is the best experimental performance value observed to date. It is at point The random variable representing the performance (following the posterior distribution given by a Gaussian process). Measured at point Conduct experiments to determine the expected performance improvement compared to the current best value. Choose to... The largest point This will serve as the next experimental point. After performing this experiment, new data will be added to the support set, and the meta-learning model will be re-adapted, updating the Gaussian process. This process will be iterated 2-3 times to efficiently explore the design space and find regions with better performance or the greatest uncertainty for validation.

[0094] (2) Uncertainty quantification modules are typically integrated into Gaussian processes. A Gaussian process provides a mean and variance for each prediction point; the variance represents the uncertainty of the prediction. During validation, in addition to examining the prediction performance values, attention is also paid to the uncertainty. For example, if the formulation... The predicted degradation rate has a mean of 6 months, but the variance is large (high uncertainty), indicating that more experiments are needed to reduce the uncertainty of the prediction. Conversely, if the predicted uncertainty of all key performance indicators is below a preset threshold (e.g., the width of the prediction interval is less than 10% of the target range), the experiment can be terminated early, the current prediction is considered reliable, and the results can be used for feedback optimization. This uncertainty-based decision-making makes the allocation of experimental resources more intelligent.

[0095] Understandably, integrating active learning (Bayesian optimization) and uncertainty quantification into the meta-learning verification framework achieves "intelligent experimental design." Instead of randomly or uniformly selecting verification points, it dynamically chooses those points most likely to improve performance or reduce cognitive blind spots. This further maximizes the efficiency of few-shot learning, rapidly constructing reliable local prediction models using the most refined experimental data. Simultaneously, uncertainty quantification provides a confidence basis for decision-making, avoiding erroneous feedback when prediction reliability is insufficient, thus improving the robustness and reliability of the entire system.

[0096] Specifically, the generation of new material formulations or gradient formulations for additive manufacturing processes also includes a multi-objective optimization process:

[0097] Multiple competitive performance objectives are defined, including biocompatibility, degradation rate, mechanical strength, printing accuracy, and material cost. The potential space of the generative adversarial network is used as the design space, and an evolutionary algorithm based on non-dominated ranking or a Pareto front optimization algorithm is used to explore it, generating a series of non-dominated solutions to form a Pareto optimal formulation set. An interactive interface is provided to allow users to specify preferences or weights, select one or more candidate formulations from the Pareto front according to preferences, and use this selection information to further constrain the conditional input of the generative adversarial network to generate formulations that better meet the user's equilibrium point.

[0098] It's important to further clarify that new material design is essentially a multi-objective optimization problem. For example, improving the mechanical strength of a material might require increasing the inorganic filler content, but this could lead to slower degradation rates and decreased printing accuracy (e.g., nozzles are more prone to clogging). Therefore, it's necessary to find a series of "Pareto optimal" solutions where no single objective can be improved without compromising at least one other objective. Specific designs include:

[0099] (1) The generator G of the conditional generative adversarial network is regarded as a mapping from the latent space Z (the space where noise z and condition c reside) to the recipe space X. In multi-objective optimization, condition c can be set to a loose objective range or be empty. Optimization is carried out in the latent space Z. Specifically, a variant of the non-dominated sorting-based genetic algorithm (NSGA-II) is adopted. First, a batch of latent vectors is generated. The corresponding recipe is obtained through generator G. Then, the performance of each formulation on multiple targets is evaluated using a performance prediction model (which can be another neural network or a version adapted from a meta-learning model) to obtain a target value vector. , where M is the number of targets.

[0100] (2) Next, the solutions are non-dominated and sorted. A solution A is said to dominate B if it is no worse than solution B on all objectives and is strictly better than solution B on at least one objective. Solutions not dominated by any other solution belong to the first non-dominated front (Pareto front). Then, the solutions on the same front are sorted according to the crowding distance. Solutions with larger crowding distances are located in sparser regions, which helps to maintain the diversity of solutions. A new population of potential vectors is generated through genetic operations such as selection, crossover, and mutation, and non-dominated sorting and selection are performed iteratively until a set of potential vectors that are approximately Pareto optimal is finally converged. These vectors are input into the generator G to obtain the corresponding Pareto optimal formula set.

[0101] (3) These Pareto optimal formulations involve performance trade-offs. The system visualizes the Pareto front as a two-dimensional or three-dimensional scatter plot through a graphical user interface (GUI) (e.g., mechanical strength as the X-axis and degradation rate as the Y-axis, with each point representing a formulation). Users can directly click on the front plot to select the formulation of interest in a region, or assign weights to different targets using sliders (e.g., mechanical strength weight 0.7, degradation rate weight 0.3). The system then automatically recommends one or more formulations that best fit the weights based on a weighted sum or a reference point-based method (e.g., TOPSIS).

[0102] (4) The user's choice constitutes new and more explicit preference information. The system encodes this preference information into a new condition vector. For example, if the user selects a formulation point in the "high strength - medium degradation" zone, This can be encoded as the average or range of performance in that region. Then, this... As a strong condition, it is input again into the conditional generative adversarial network to run the generation process once more. Because the conditions are stronger and more specific, the generated recipes will be densely distributed near the Pareto front region selected by the user, providing the user with richer and more nuanced candidate solutions under this preference, realizing a cycle of "user interaction - preference learning - targeted generation".

[0103] Understandably, by integrating multi-objective optimization, this invention transforms single, vague performance targets into a systematic exploration and visualization of the trade-off space. This not only helps materials scientists more comprehensively understand the inherent conflicts and connections between different performance indicators but also returns design control to the user. Users can make final trade-off decisions based on actual application scenarios (e.g., load-bearing bone defects require higher strength, while non-load-bearing areas may be more concerned with degradation matching). The system then performs targeted refinement design based on these decisions, ensuring that the final output highly matches the user's actual needs, achieving intelligent design through human-machine collaboration.

[0104] Specifically, this also includes integrating the generated formula with the additive manufacturing equipment control system:

[0105] A dedicated conversion module converts formula data and its associated process parameters into control instructions executable by the equipment, including G-code, slice files, or direct drive signals. During the formula generation stage, an integrated equipment hardware constraint model is used, which includes nozzle diameter, heating temperature range, platform motion accuracy, and material delivery system limitations, ensuring that the generated formula is executable within the physical limits of the equipment. A closed-loop control process is established, and virtual simulation verification is performed during the printing preparation stage. The formula or process parameters are then fine-tuned based on the simulation results.

[0106] It's important to further explain that the seamless integration from digital recipes to physical entities is crucial for achieving intelligent manufacturing. A dedicated conversion module acts as a "translator." Its input is a structured recipe data object, containing a list of material components, gradient functions (if applicable), and a set of process parameters. Its output is executable instructions for a specific model of additive manufacturing equipment (such as a certain brand of dual-nozzle bio-3D printer).

[0107] (1) For homogeneous material formulations, the conversion process is relatively straightforward. For example, for a formulation with material composition of [material A: 70%, material B: 30%], if the equipment is equipped with a dual-feed system, the conversion module will generate instructions to control the speed ratio of the two feed motors to achieve an accurate volume mixing ratio. At the same time, the process parameters in the formulation (such as printing temperature 220°C, printing speed 40mm / s, layer thickness 0.15mm) will be directly written into the parameter settings of the slicing software, ultimately generating a standard G-code file.

[0108] (2) For gradient material formulations, the conversion process is more complex. Assume the formulation specifies that the volume fraction of material A linearly gradients along the Z-axis from 0% at the bottom to 100% at the top. The conversion module needs to interact deeply with the slicing engine. First, the 3D model is sliced ​​according to the set layer thickness. Then, for each layer, the theoretical volume ratio of materials A and B required for that layer is calculated based on its Z-axis height. Next, when generating the fill path (G-code) for that layer, the extrusion rate of the dual nozzles is dynamically adjusted to achieve a smooth transition of material proportions within each layer (if needed) and between layers. This may require generating special G-code instructions that support real-time extrusion rate adjustment (such as dynamic adjustment based on the E-value).

[0109] (3) To ensure the executability of the generated recipes, a device hardware constraint model is integrated within the design loop of the generative adversarial network. This model is a set of rules or a lightweight discriminant network. It is invoked immediately after the generator G outputs a candidate recipe to check whether the parameters in the recipe exceed the physical capabilities of the current target device. For example, the rules may include: if the recommended printing speed > the maximum movement speed of the device, then mark it as a violation; if the recommended nozzle temperature < the melting point of material A or > the maximum safe temperature of the device, then mark it as a violation; if the viscosity ratio of the two materials exceeds 10:1, which may lead to uneven blending or clogging, then mark it as a violation. Once marked as a violation, the recipe will be assigned a very low "manufacturability score" and will be eliminated in subsequent optimization screening, or directly fed back to the generator to adjust its output.

[0110] (4) The closed-loop control process is triggered after virtual simulation verification. As mentioned earlier, the physics-based simulation predicts the printing process. If the simulation finds potential defects (such as slight warping), the closed-loop control process will not directly reject the recipe, but will instead initiate a "parameter fine-tuning" subprocess. For example, if the simulation predicts warping, the fine-tuning subprocess may automatically try to increase the print platen temperature by 5-10°C or reduce the contour printing speed by 20%, and then rerun the fast simulation. If the adjusted simulation results meet the requirements, the parameters in the original recipe are replaced with the fine-tuned process parameters, and then sent to the conversion module. In this way, manufacturability issues are resolved in the virtual world as much as possible before actual printing.

[0111] Understandably, by integrating a dedicated conversion module and a hardware constraint model, this invention establishes a seamless digital thread from material design to equipment execution, ensuring the feasibility of the design outcomes. Early intervention with the hardware constraint model prevents the design of "unprintable" formulations, saving subsequent iteration time. Furthermore, the closed-loop control process, including virtual simulation and parameter fine-tuning, is equivalent to adding a high-fidelity "digital twin" testing step before physical manufacturing. This allows for early prediction and correction of problems, significantly improving the success rate of the first print, reducing material waste and machine downtime, and ensuring efficient, reliable, and intelligent manufacturing.

[0112] Specifically, the closed-loop control process also includes real-time online adjustment:

[0113] In the additive manufacturing printing process, sensors integrated into the equipment are used to monitor key process variables, including extrusion pressure, temperature distribution, layer thickness deviation, and degree of curing. The real-time data stream is input into a pre-trained meta-learning model, which dynamically evaluates the deviation between the current printing state and the expected state and generates adjustment signals. The adjustment signals are fed back to the online fine-tuning module of the generative adversarial network model to adjust the material formulation ratio or process parameter settings of subsequent printing layers in real time, so as to achieve adaptive printing and performance compensation and ensure the consistency of the final part performance.

[0114] It should be further explained that real-time online adjustment is the ultimate means of dealing with uncertainties and disturbances in the manufacturing process, forming a rapid closed loop of "online perception-decision-execution". This closed loop is independent of the aforementioned offline design closed loop, but shares some models with it. The specific design includes:

[0115] (1) The sensing system is the foundation of real-time closed-loop processing. A pressure sensor is installed near the printhead to monitor the melt extrusion pressure; an abnormal increase in pressure may indicate nozzle blockage or a sudden change in material viscosity. An infrared thermal imager or thermocouple monitors the temperature field of the deposited filament; insufficient temperature will affect interlayer bonding. A laser displacement sensor or line laser scanner monitors the surface height (layer thickness) of the printed layer to detect uneven spreading or warping. For photopolymerization processes, an ultraviolet light intensity sensor may be used to monitor the exposure energy. These sensors acquire data at high frequencies (e.g., 10-100 Hz) to form a real-time data stream.

[0116] (2) The pre-trained meta-learning model acts as an "online state evaluator" here. During the meta-training phase, this model learns the mapping relationship from the process variable sequence to the print quality / performance deviation. During the printing process, it takes sensor data (such as the pressure and temperature sequence of the last 10 seconds) within the current and recent time window as input. Due to its rapid adaptability, it can combine the specific recipe of the current printing task (as conditional information) to output an evaluation of the current printing state in real time, such as predicting the deviation between the actual mechanical properties of the current layer and the target value. Or diagnose the type of problem that is currently occurring (such as "high risk of poor interlayer bonding").

[0117] (3) The online fine-tuning module receives the deviation signal or diagnostic results output by the state evaluator. This module is essentially a lightweight, fast-response version of the generator G in a generative adversarial network. Instead of regenerating the entire recipe, it calculates the adjustment amount for key parameters of subsequent printing layers (such as the extrusion ratio of material A, printing speed, and nozzle temperature of the next layer) based on the current deviation. This calculation can be based on a pre-trained policy network that takes state biases as input and parameter adjustments as output. It is immediately sent to the device controller to modify subsequent G-code instructions or directly control the actuator.

[0118] (4) A concrete example: Printing a bone scaffold with gradient mechanical properties. The goal is to gradually decrease stiffness from bottom to top. When printing to the middle, the pressure sensor shows that the pressure is slightly lower than expected, and the infrared thermometer also shows that the temperature is slightly lower. The state evaluator combines this information and determines that the actual stiffness of the current layer may be lower than the value on the expected gradient curve. The online fine-tuning module receives the "stiffness is lower than expected" signal and, according to a predetermined strategy, decides to temporarily increase the proportion of high-modulus material (e.g., increase by 2%) and slightly increase the printing temperature (e.g., increase by 5°C) in several subsequent layers to compensate. These adjustments are made in real time and dynamically to ensure that the overall performance gradient of the final product is as close as possible to the design goal.

[0119] Understandably, real-time online adjustments upgrade the manufacturing process from "open-loop execution of preset programs" to "closed-loop adaptive manufacturing." It can sense fluctuations in the manufacturing process in real time (such as batch differences in materials, changes in environmental temperature and humidity, and minor equipment wear) and make timely compensatory adjustments. This gives the manufacturing system anti-interference capabilities and robustness, ensuring that even under less than ideal conditions, the performance of the final part remains highly consistent with design expectations. This significantly improves product consistency and reliability, and is of great value for the manufacturing of medical implants, which have extremely high requirements for performance consistency.

[0120] Specifically, this also includes establishing a continuous iterative optimization loop:

[0121] After each experimental verification or actual printing of the generated formula, the obtained performance and process data are collected and labeled, and expanded into the historical training database. The generative adversarial network model and meta-learning model are retrained periodically using the expanded data, and their parameters are updated to absorb new knowledge. This cycle enables the system to adapt to the introduction of new material systems, changes in process parameters, or changes in performance targets, gradually improving the first success rate of formula generation and expanding its designable material space, forming a self-evolving generative design system for new material formulas.

[0122] It's important to further clarify that a truly intelligent system should possess the ability to continuously learn from experience. The continuous iterative optimization loop established by this system forms the foundational architecture for achieving this capability. This loop is not a simple accumulation of data, but a systematic knowledge growth engine, specifically designed as follows:

[0123] (1) Data collection and labeling are the starting point of the cycle. Every experimental verification, whether it is a small-scale performance test (such as a compression test or a degradation experiment) or a complete printing process, will generate valuable data. This data includes:

[0124] a) Input data: The final formula used (ingredients, process parameters).

[0125] b) Process data: All sensor data and environmental data recorded during the experiment or printing process.

[0126] c) Output Data: The final measured performance indicators and the quality assessment results of the fabricated parts (such as porosity shown by micro-CT and microstructure shown by scanning electron microscopy). This data is collected, cleaned, aligned, and timestamped and labeled with experimental context tags in a standardized structured format before being stored in a central database. A high-quality labeling system is crucial; for example, labels should include information such as the material system category, equipment model, and operator.

[0127] (2) Periodic retraining of the model is a process of knowledge absorption. The model retraining process is triggered when the newly accumulated data reaches a certain scale (e.g., 50 new valid data points are added), or when a completely new material system is introduced (e.g., when research on biodegradable magnesium alloys begins). Retraining does not start from scratch, but rather uses the current model parameters as pre-training weights and fine-tunes them using a mixture of old and new datasets. For generative adversarial networks, adversarial training continues with both old and new data to ensure that the generated distribution can cover the new material system. For meta-learning models, new data is constructed into new meta-tasks and added to the meta-training task distribution for re-meta-optimization, enabling the model to adapt to new tasks.

[0128] (3) The frequency and strategy of retraining need to be balanced. One strategy is to set a fixed data volume threshold (e.g., every 100 new data points). Another more complex strategy is to actively monitor the degradation of model performance. For example, when the failure rate (e.g., failure to meet key performance indicators) of the recipes generated by the system in experimental verification exceeds a certain threshold (e.g., 20%), retraining is triggered. During retraining, newer data can be given slightly higher weights to allow the model to adapt to the latest process or material changes more quickly.

[0129] (4) Through this continuous cycle, the system achieves self-evolution. Initially, the system may only be good at designing formulations based on a few common biopolymers. As the cycle progresses, it gradually learns about more diverse materials (such as synthetic polymers, natural polymers, ceramics, and metals) and their composite rules. It also gradually adapts to the characteristics of different printer models. Ultimately, the system becomes an experienced "AI material designer" covering a wide range of material systems and processes. The probability of its generated formulations being experimentally verified or successfully printed for the first time will become higher and higher, and the cycle of designing new formulations will become shorter and shorter, forming a powerful and continuously growing design-manufacturing intelligent agent.

[0130] Understandably, establishing a continuous iterative optimization cycle is key to transforming this invention system from a "tool" into a "partner." It breaks down the static knowledge boundaries of traditional design software, enabling the system to continuously accumulate experience, improve itself, and expand during use. This means the system's value will increase over time and with increased usage. By deploying and using this system, enterprises and research institutions not only accelerate the development of individual projects but also continuously build and enrich a dedicated, digital "materials design and manufacturing knowledge base," constituting a long-term core competitiveness in the field of additive manufacturing new materials development. This system forms a self-evolving ecosystem capable of adapting to change and learning and growing.

[0131] Furthermore, the implementation process of this method is further illustrated below through a specific operational case. The objective of this case is to design a composite scaffold formulation for skull defect repair, which needs to be manufactured using a fused deposition modeling process and meet the following target performance requirements:

[0132] (1) The compression modulus is between 100MPa and 150MPa to match the mechanical properties of cancellous bone;

[0133] (2) In simulated body fluids, the rate of mass loss (degradation rate) within 12 weeks was between 20% and 35% to support new bone ingrowth;

[0134] (3) Cell survival rate >90% to ensure biocompatibility.

[0135] Furthermore, specific operational examples are as follows:

[0136] 1. Data preparation and model loading:

[0137] The system accesses a historical database containing 200 sets of material formulations, process parameters, and performance data based on polylactic acid (PLA), polycaprolactone (PCL), and hydroxyapatite (HA) systems. The data has been normalized. Pre-trained Conditional Generative Adversarial Network (WGAN-GP architecture) and Model-Independent Meta-Learning (MAML) models are loaded. The generator for the generative adversarial network is then used. and discriminator The parameters have converged. Meta-learning model Meta-training has been completed on multiple material property prediction tasks, yielding good initial parameters. .

[0138] 2. Setting goals and generating initial recipes:

[0139] The user inputs the above performance targets through the interface. The system encodes the targets as condition vectors. The first two elements represent the lower and upper limits of compressive modulus (MPa), the middle two represent the lower and upper limits of degradation rate (12-week mass loss rate), and the last represents the lower limit of cell viability.

[0140] Generators of Generative Adversarial Networks Receive condition vector and a random noise vector sampled from a standard normal distribution. (For example, ), generate the first candidate recipe . Output as structured vectors, for example: .

[0141] 3. Simulation-based initial screening for manufacturability:

[0142] formula The material was fed into a physics-based simulation module. The simulation used computational fluid dynamics to model the melt flow within a nozzle (0.4 mm in diameter). The simulation calculated an extrusion pressure of 4.8 MPa at 205°C and the material's viscosity, below the equipment limit (10 MPa). Simultaneously, the thermal deformation simulation predicted a warpage of 0.15 mm, within acceptable limits (<0.3 mm). Manufacturability score. It was calculated as 0.92 (out of 1). Since the score was above the threshold (0.7), the recipe... After passing the initial screening, the experiment will proceed to the verification stage.

[0143] 4. Limited experimental verification and feedback on meta-learning guidance:

[0144] The system initiates an active learning loop. The initial support set is empty. Meta-learning model. Based on its prior knowledge, predict The performance was as follows: compressive modulus 135 MPa, degradation rate 28%, and cell viability 92%. Uncertainty quantification showed that the standard deviation of the degradation rate prediction was relatively large (±5%).

[0145] To reduce uncertainty and validate predictions, Bayesian optimization of the acquisition function suggests... and its two perturbation variants (HA: +2%) (Printing speed: -5mm / s) Conduct experiments. Perform these three experiments to obtain real data: .

[0146] Will As a support set, the meta-learning model is quickly adapted. The adaptation process performs a three-step gradient descent, with a learning rate of... Adapted model Re-prediction Its uncertainty is significantly reduced. Model evaluation Deviation from target: The degradation rate (31%) is close to the upper limit (35%), posing a risk of exceeding the target. Calculate the deviation loss. (Assuming a threshold).

[0147] This deviation loss Backpropagation to the generator of the generative adversarial network Calculate the gradient And update its parameters This will guide the next round of formulations to focus more on reducing the degradation rate.

[0148] 5. Multi-objective optimization and decision-making:

[0149] Upon receiving feedback, the system runs a multi-objective optimization process. The NSGA-II algorithm is used to explore the latent space of the generative adversarial network, generating a Pareto front containing 15 non-dominated solutions. The front end displays the trade-off between modulus and degradation rate.

[0150] Users observe the forefront through the interface and discover a formula. (PLA: 68%, PCL: 20%, HA: 12%, temperature: 200°C, degradation rate: 30mm / s) This formulation achieves a good balance between modulus (118MPa) and degradation rate (26%). The user selected this formulation.

[0151] The system encodes this preference into a new condition vector. And run the condition generation again, in Five refined candidate recipes were generated nearby.

[0152] 6. Closed-loop verification and final output:

[0153] The meta-learning validation in step 4 was repeated for the five refinement candidates (two experiments for each). Finally, the formulation... =[PLA: 69.5%, PCL: 18.2%, HA: 12.3%, Printing temperature: 198°C, Printing speed: 32mm / s, Layer thickness: 0.20mm] won with the highest overall score. Its verified performance is: compressive modulus 122MPa, degradation rate 28%, cell viability 93%, fully meeting all targets.

[0154] The system calls the conversion module to... The parameters are converted into G-code for a specific brand of dual-nozzle 3D printer. The simulation module performs virtual verification of the final printing process and passes the verification.

[0155] formula The performance prediction report, process parameter file, and G-code are packaged and output to the user for actual printing.

[0156] 7. Data Feedback and Model Updates:

[0157] After the actual printing is completed and the part is more comprehensively characterized (such as micro-CT, mechanical testing), the new data obtained (including process sensor data) is labeled and added to the historical database.

[0158] Once 100 sets of such new data have been accumulated, the system automatically triggers the incremental training process of generative adversarial networks and meta-learning models, using a mixture of new and old datasets to fine-tune the model parameters and complete a knowledge iteration.

[0159] Understandably, this case study demonstrates the complete process of the method from setting objectives to outputting a manufacturable recipe. Through a generative adversarial network, the system learns from historical data and proactively generates initial candidate recipes that satisfy multi-objective constraints without human intervention. The meta-learning model efficiently completes the verification and performance evaluation of candidate recipes in just 5 experiments (3 initial + 2 refinements), and incorporates feedback into design optimization. The multi-objective optimization and user interaction stages clarify performance trade-offs and achieve human-machine collaborative decision-making. The final output recipe... Not only does it meet all preset biological and mechanical performance targets, but it also passes manufacturability simulation verification and can directly drive a printer. The entire process compresses the dozens or even hundreds of "trial and error" experiments that may be required in traditional materials research and development into a single-digit number of "intelligent" experiments, and automatically generates manufacturing instructions, significantly accelerating the process from design to manufacturing. All computational steps in the case (such as gradient descent, expected improvement calculation, and Pareto sorting) are based on the algorithms described in the aforementioned specific implementation, verifying their feasibility and effectiveness.

[0160] In a preferred embodiment, this application also provides an electronic device, the electronic device comprising:

[0161] The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, implement the described generative design method for novel material formulations for additive manufacturing. The computer device can be broadly categorized as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the method of the present invention.

[0162] This invention can be implemented as a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the methods of embodiments of the invention to be performed. In one embodiment, the computer program is distributed across multiple network-coupled computer devices or processors, such that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be executed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be executed by one or more computer devices or processors, and one or more other method steps / operations may be executed by one or more other computer devices or processors. One or more computer devices or processors may execute a single method step / operation, or execute two or more method steps / operations.

[0163] Those skilled in the art will understand that the method steps of this invention can be performed by a computer program instructing related hardware, such as a computer device or processor, to perform the steps of this invention when executed. Depending on the context, any references herein to memory, storage, databases, or other media may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0164] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0165] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A new material recipe generation design method for additive manufacturing, characterized by, include: S1. Utilize generative adversarial network models to learn the mapping relationship between the composition, process parameters, and performance indicators of known biomaterials; S2. Generate new material formulations or gradient formulations for additive manufacturing processes based on the mapping relationship; S3. The generated formula is validated and fine-tuned using a small amount of experimental data through a meta-learning model.

2. The method of claim 1, wherein, The method of using a generative adversarial network model to learn the mapping relationship between the composition, process parameters, and performance indicators of known biomaterials includes: A training dataset was constructed using known biomaterial composition data, process parameter data, and performance index data. The process parameter data included at least one of the printing temperature, printing speed, layer thickness, and curing parameters in the additive manufacturing process. A conditional generative adversarial network architecture was adopted, in which the target performance index was simultaneously input into the generator and discriminator as a conditional vector. This enabled the generator to output candidate formulations that met the target performance index, while the discriminator evaluated the consistency between the generated formulations and the real formulations under given conditions. The network weights of the generator and discriminator were iteratively optimized through an adversarial training process until the generator could stably output formulations that met the performance constraints. This training process integrated a gradient penalty loss function to improve training stability.

3. The method according to claim 2, characterized in that, The generator adopts an encoder-decoder structure, where the encoder maps the component data and process parameter data into a low-dimensional latent space representation, and the decoder reconstructs or generates a new formula from this latent space representation. A multi-head self-attention mechanism is embedded in the encoder-decoder to capture cross-modal dependencies between components and process parameters and to focus on the impact of key features on performance metrics. During training, spectral normalization is introduced to constrain the weight matrix of the discriminator, and a mini-batch discrimination mechanism is combined to increase the diversity of generated recipes.

4. The method according to claim 3, characterized in that, The decoder outputs a formula including the mixing ratio of various bioactive materials, gradient distribution function, and recommended additive manufacturing process parameter adjustment values. The formula is represented in structured vector or matrix form and can be directly converted into an input format that the additive manufacturing equipment can parse. After generating the formula, a physics-based simulation model is integrated to verify its manufacturability. The simulation model simulates the flow, solidification, and interlayer bonding behavior of materials during the printing process, predicts possible porosity, cracks, or deformation defects, and feeds the simulation results back to the generator to adjust the formula generation strategy.

5. The method according to claim 1, characterized in that, The process of validating and fine-tuning the generated formula using a meta-learning model with a small amount of experimental data includes: A model-independent meta-learning framework is adopted to perform meta-training on multiple heterogeneous material formulation verification tasks to learn an initial model parameter that adapts quickly. For each newly generated formulation, a small number of key experiments are designed and executed to obtain its performance data, forming a support set. The meta-learning model is quickly adapted using the support set data through a small number of gradient descent steps to obtain a task-specific model. This model is used to predict the unmeasured performance of the formulation or evaluate the deviation from the target performance. The loss gradient is calculated based on the deviation signal to fine-tune the parameters of the generative adversarial network model.

6. The method according to claim 5, characterized in that, The meta-training process of the model-independent meta-learning framework includes: The model samples support and query sets across multiple tasks. On the support set, it computes task-specific losses and updates model parameters. On the query set, it evaluates the meta-loss and computes meta-gradients with respect to initial parameters. The meta-optimizer updates the initial parameters to enable the model to adapt quickly to new tasks. During the validation phase, it actively selects experimental points using a Bayesian optimization strategy to maximize expected improvement or minimize uncertainty, thereby efficiently acquiring the most informative experimental data for model adaptation. The meta-learning model also integrates an uncertainty quantification module to provide confidence estimates for performance prediction, guiding early cessation or further exploration in the validation process.

7. The method according to claim 1, characterized in that, The generation of new material formulations or gradient formulations for additive manufacturing processes also includes a multi-objective optimization process: Multiple competitive performance objectives are defined, including biocompatibility, degradation rate, mechanical strength, printing accuracy, and material cost. The potential space of the generative adversarial network is used as the design space, and an evolutionary algorithm based on non-dominated ranking or a Pareto front optimization algorithm is used to explore it, generating a series of non-dominated solutions to form a Pareto optimal formulation set. An interactive interface is provided to allow users to specify preferences or weights, select one or more candidate formulations from the Pareto front according to preferences, and use this selection information to further constrain the conditional input of the generative adversarial network to generate formulations that better meet the user's equilibrium point.

8. The method according to claim 1, characterized in that, This also includes integrating the generated recipe with the additive manufacturing equipment control system: A dedicated conversion module converts the formula data and its associated process parameters into control instructions that the equipment can execute, including G-code, slice files, or direct drive signals. During the formula generation stage, an integrated equipment hardware constraint model is used, which includes nozzle diameter, heating temperature range, platform motion accuracy, and material delivery system limitations, to ensure that the generated formula is executable within the physical limits of the equipment. Establish a closed-loop control process, conduct virtual simulation verification during the printing preparation stage, and fine-tune the formula or process parameters based on the simulation results.

9. The method according to claim 8, characterized in that, The closed-loop control process also includes real-time online adjustment: During the additive manufacturing printing process, sensors integrated into the equipment are used to monitor key process variables, including extrusion pressure, temperature distribution, layer thickness deviation, and degree of curing. The real-time data stream monitored is input into a pre-trained meta-learning model, which dynamically evaluates the deviation between the current printing state and the expected state and generates adjustment signals. The adjustment signal is fed back to the online fine-tuning module of the generative adversarial network model to adjust the material formula ratio or process parameter settings of subsequent printing layers in real time, so as to achieve adaptive printing and performance compensation and ensure the consistency of the final part performance.

10. The method according to claim 9, characterized in that, This also includes establishing a continuous iterative optimization loop: After each experimental verification or actual printing of the generated formula, the obtained performance and process data are collected, labeled, and expanded into the historical training database. Regularly retrain the generative adversarial network model and the meta-learning model using the augmented data, and update their parameters to absorb new knowledge; This cycle enables the system to adapt to the introduction of new material systems, changes in process parameters, or changes in performance targets, gradually improving the first-time success rate of formulation generation and expanding its designable material space, forming a self-evolving new material formulation generation design system.

Images