3.1 Design of AI Driven Evaluation Model For 3D Printing Technology in Relief Craft

This section first constructs an evaluation index system for 3D printing technology in relief technology, including accuracy, speed, product quality, and AI integration level. Then, the Q-learning algorithm and swarm intelligence are used to optimize the judgment matrix and dynamically adjust the weights. Finally, a comprehensive evaluation is conducted by combining Q-learning-based combination weight algorithm and ideal point method. The proposed model can quantitatively evaluate the accuracy, efficiency, and quality of 3D printing technology in relief technology.

Faced with the quality control challenges and efficiency issues in the development of 3D printing technology for relief craft, as well as the increasing demand for personalized products, it is particularly important to build a scientific and reasonable evaluation index system for 3D printing technology for relief craft ^[17,18]. Therefore, this study starts from the entire process of 3D printing technology and constructs a set of evaluation index system for 3D printing technology of relief craft. This index system has six characteristics: system integrity, layering, relative independence, guidance, practicality, and comparability. Considering the current situation of 3D printing technology in relief craft, 11 representative and covering evaluation indicators are selected from three dimensions: design accuracy, printing speed, and product quality. Through this indicator system, the application level of 3D printing technology in relief technology can be effectively evaluated, as indicated in Table 2.

In Table 2, the development level indicators of relief 3D printing technology include relief printing accuracy ratio (k1), 3D printing speed (k2), printing accuracy improvement rate (k3), and relief printing path optimization degree (k4). In terms of product quality, evaluation indicators include qualified yield (s1), surface smoothness (s2), and material utilization rate (s3). Regarding the AI integration level, indicators include AI optimization ratio (f1), Q-learning algorithm optimization effect (f2), automated printing ratio (f3), and intelligent path planning completion degree (f4). These indicators possess system integrity, layering, and directionality, which can effectively evaluate the practical application level of 3D printing technology in relief technology.

To efficiently estimate the application progress of 3D printing technology in relief craft, an AI evaluation model based on Q-learning algorithm optimization is designed. This model combines AI driven learning mechanisms and path planning algorithms, adjusts subjective and objective parameters, and combines traditional evaluation methods to assign weights and comprehensively analyze selected evaluation indicators, achieving quantitative and comprehensive evaluation of the application level of 3D printing technology in relief craft. The hierarchical analysis structure is shown in Fig. 1.

In Fig. 1, the AI analytic hierarchy process (AI-AHP) based on Q-learning algorithm is utilized to calculate the subjective weights of the evaluation indicators for 3D printing technology in relief craft. The specific steps include building a hierarchical structure model, and establishing a judgment matrix, using Q-learning algorithm to optimize weights and performing consistency testing to obtain the total weight. As a global optimization method, Q-learning algorithm can optimize the judgment matrix, avoid possible inconsistency issues, and improve the accuracy of evaluation index weight calculation ^[19,20]. The judgment matrix is shown in Eq. (1).

In Eq. (1), $n$ means the amount of evaluation indicators and $a_{nn} $ denotes the comparison of importance between indicators. The relative weight of elements under consistency testing is shown in Eq. (2).

To optimize the consistency of the evaluation process, Q-learning is used to adjust the judgment matrix. The consistency check formula is shown in Eq. (3).

In Eq. (3), $CR$ means the consistency ratio coefficient, $\lambda _{\max } $ refers to the maximum eigenvalue, and $RI$ expresses the random consistency index. The introduced Q-learning algorithm improves the consistency of judgment matrix through learning and iterative optimization, thereby enhancing the accuracy and efficiency of 3D printing technology evaluation in relief craft. The timing optimization and path planning functions of AI further enhance the dynamic adaptability of the evaluation model, making it more in line with the application requirements of 3D printing technology in relief craft. The consistency index values are indicated in Table 3.

In Table 3, Data 1 and Data 2 represent the consistency indices obtained by introducing Q-learning algorithm optimization under different judgment matrix orders. Among them, Data 1 is the consistency index of Q-learning initial parameter optimization, mainly referring to the consistency index value of the optimized judgment matrix under the initial conditions or specific parameter settings of the Q-learning algorithm. Data 2 shows the consistency index values of the optimized judgment matrix for different Q-learning parameters under another set of parameter settings in the Q-learning algorithm. On this basis, a nonlinear model can be obtained from the consistency judgment matrix and group information, as shown in Eq. (4).

Q-learning algorithm and swarm intelligence are used to construct a nonlinear model, optimize the judgment matrix of 3D printing relief craft evaluation indicators, and dynamically adjust weights. The QLEM method utilizes information entropy to quantitatively evaluate indicator weights, and Q-learning further optimizes these weights in a temporal manner, improving the objectivity and adaptability of the evaluation. The key steps include establishing a standardized indicator matrix, calculating feature weights, obtaining entropy weights, and finally calculating the coefficient of difference. The indicator weight is shown in Eq. (5).

In Eq. (5), $k_{j} $ represents the entropy weight. By combining the temporal optimization ability in AI technology, dynamic adjustment of indicator weights can be achieved.

After calculating the subjective and objective weights, a combination weight algorithm based on Q-learning is used to integrate the two weights and obtain the comprehensive weight of a single indicator. The goal of this combination weight algorithm is to find a weight combination that minimizes the information difference between subjective and objective weights while maintaining the dynamic adaptability of the evaluation model, to make the evaluation results more accurate and reasonable. The combined weight is shown in Eq. (6).

In Eq. (6), $j$ refers to the number of indicator items, $w_{j} $ means the combination weight, $w_{1j} $ means the subjective weight, and $w_{2j} $ means the objective weight. Finally, the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) is utilized to conduct a comprehensive evaluation of 3D printing technology for relief craft. The overall model structure is shown in Fig. 2.

This method judges the quality of an object by evaluating its relative proximity to positive and negative ideal solutions. The ideal point method can fully utilize evaluation indicator data, reduce information loss, and improve the reliability of evaluation results. The operation steps are as follows: establishing a decision matrix, performing dimensionless data processing, calculating the weighted normalized decision matrix, obtaining positive and negative ideal solutions, and calculating the distance between the evaluation scheme and the positive and negative ideal solutions. The positive ideal solution is shown in Eq. (7).

In Eq. (7), $z_{ij} $ represents the weighted data. The negative ideal solution is denoted in Eq. (8).

By combining the indicator weights obtained from the Q-learning algorithm-based AI evaluation method with the evaluation results calculated by the TOPSIS method, an analysis is conducted in a non-integrated evaluation model. This model can quantitatively evaluate the accuracy, efficiency, and product quality of 3D printing technology in relief technology, thereby promoting the application of 3D printing technology in relief craft. The distance from the object to the positive ideal solution is shown in Eq. (9).

The distance from the object to the negative ideal solution is shown in Eq. (10).

The comprehensive evaluation value is indicated in Eq. (11).

This model has broad application prospects, which can be utilized to assess the performance of different 3D printers in the application of relief craft, providing support for the selection and optimization of 3D printing technology in relief craft.

3.2 Design of 3D Printing Path Planning Optimization Model Based on Q-Learning Algorithm

After completing the AI-driven evaluation model for 3D printing technology in relief technology, a 3D printing path planning optimization model based on Q-learning algorithm is constructed. Firstly, based on the principles of path planning timing optimization, the model hypothesis is established and an objective function is established using the principles of accuracy, efficiency, cost, and stability. At the same time, the balance constraint between printing accuracy and speed are set. Subsequently, under model constraints, the Q-learning algorithm is optimized and combined with multi-objective decision-making techniques to address the challenges of 3D printing path planning. Finally, the evaluation model integrates AI experience replay, selects the global optimal path through standardized objective function, local optimal path search and comparison, and formulates the optimal path plan. Under the comprehensive influence of numerous influencing factors, a 3D printing path planning optimization model based on Q-learning algorithm is constructed. Firstly, based on the principle of path planning timing optimization, the following model assumptions are established as shown in Fig. 3.

According to the optimization principles, the study adopts the principles of accuracy, efficiency, cost, and stability to establish the objective function. The accuracy principle is to maximize printing accuracy and ensure that the details of the relief craft can be accurately presented. The efficiency principle is to minimize the total printing time and improve the efficiency of 3D printing operations. The cost principle is manifested in reducing material consumption and maintenance cost during the 3D printing process. The stability principle is manifested in ensuring the stable operation of the 3D printing process, in order to reduce the risk of printing failure. In terms of model constraints, the balance between printing accuracy and speed is taken as the priority constraint.

The printing accuracy constraint is shown in Eq. (12).

In Eq. (12), $D_{t} $ represents the 3D printing project of Class $t$ relief craft, and $\varepsilon _{j,k,t} $ represents the accuracy standard required for the printing task of Class $j$ and $k$ projects in the $t$th printing task. The printing efficiency constraint needs to satisfy Eq. (13).

In Eq. (13), $L_{\max ,t}^{D} $ represents the maximum printing task quantity, $H_{j,k,t} $ represents the printing time, and $r$ represents the printing efficiency. The material consumption constraint needs to satisfy Eq. (14).

In Eq. (14), $P_{j,t} $ means the total material consumption, and $P_{j,t,\max } $ means the maximum allowable material consumption. The stability constraint for printing needs to meet Eq. (15).

In Eq. (15), $\lambda _{j} $ means the stability index of the task, $p_{j} $ means the upper limit of stability decline rate, $E_{j,t} $ means the amount of printing tasks, and $z_{j} $ means the upper limit of stability increase rate. The research aims to address the challenges of 3D printing path planning by optimizing the Q-learning algorithm and combining multi-objective decision-making techniques. In the optimization process of 3D printing technology in relief technology, a quantitative evaluation objective function is established to accurately measure the material saving effect. According to the material consumption constraint, the objective function can be defined, as shown in Eq. (16).

In Eq. (16), $F_{m} $ represents the objective function of material utilization savings. $N$ represents the total number of printing tasks. $M_{i} $ represents the actual material consumption of the $i$-th printing task. $M_{\max } $ represents the maximum allowable material consumption. This objective function accumulates the ratio of the actual material consumption $M_{i} $ to the maximum allowable consumption $M_{\max } $ for all printing tasks to form a comprehensive indicator, thereby quantifying the material saving effect. The optimization goal is to minimize the total material consumption ratio while meeting the process requirements, in order to achieve the optimization principle of cost reduction. Abandoning traditional path strategies and incorporating AI experience replay to enable learning to occur on specific trajectories, ensuring the integrity, robustness, and non-redundancy of action reward encoding. The encoding reflects the path selection and its impact on printing accuracy and efficiency, and adjusts algorithm parameters such as learning rate, discount factor, and exploration rate. The algorithm process is shown in Fig. 4.

In Fig. 4, the evaluation model incorporates AI experience replay to ensure that learning takes place on specific trajectories, making the action reward encoding complete and robust. By standardizing the objective function, the dimensions of printing accuracy, speed, and material consumption are eliminated to unify the evaluation system. After the initial strategy is generated, the strategy with the highest reward value is directly calculated and retained. The learning update rules of the Q-learning and $\varepsilon$-greedy algorithm are applied to local optimal path search. Before reaching the predetermined number of learning times, continuous exploration and optimization are carried out. Finally, all local optimal solutions are compared to select the global optimal path and formulate the best path plan. This model significantly improves the accuracy and efficiency of path planning by improving various parameters such as learning rate, discount factor, and exploration rate.

4.1 Analysis of Q-Learning Algorithm Optimization Effect in Evaluation Model

To ensure the accuracy of the experimental outcomes, the research is conducted on the testing and optimization analysis of the Q-learning algorithm using a cloud server platform. In the 3D printing technology of relief craft, Q-learning algorithm is used for path planning. Through a cloud server platform, an AI optimized path planning environment is simulated during the 3D printing process. The Q-learning algorithm is carefully adjusted and tested for the learning and decision-making processes of different relief patterns. Key algorithm parameters include reward function, learning rate, and discount factor. The hardware configuration includes optimizing the processor, memory, and storage space of the cloud server to ensure efficient data processing. Machine learning frameworks and programming environments are selected for software configuration. The details of software and hardware configuration and model parameter settings are shown in Table 4. Table 4 contains the software and hardware configurations and model parameter settings for the cloud server platform used for Q-learning algorithm testing and optimization analysis. The hardware configuration includes key components such as high-performance processors and high-capacity memory to ensure efficient data processing of 3D printing embossing technology in the Q-learning algorithm path planning process. The software configuration selects an adapted programming environment and machine learning framework. In addition, the parameter adjustment of the Q-learning algorithm is also set to meet the complex process requirements of 3D printing.

To conduct performance testing on the evaluation model, the average convergence of Q-learning algorithm in the application of relief craft path planning in 3D printing technology was tested, which was also compared with Improved Heuristic-based Search Algorithm (IHSA) and Ant Colony Optimization for Path Planning (ACOPP) for comparative testing. The average fitness value of the model was used as the indicator for testing. The test results are expressed in Fig. 5.

In Fig. 5, the Q-learning algorithm optimization model used in the study exhibited fast convergence speed and excellent average fitness. The optimal average fitness value of the model reached $1.519\times10^{-3}$. The results indicates that compared with other comparative models, the Q-learning algorithm has significantly improved optimization speed and efficiency.

According to the research on 3D printing technology path planning in the predicted relief craft by 2030, if no optimization measures were taken, the efficiency and accuracy of path planning would show a downward trend, as shown in Fig. 6.

In Fig. 6, the parameter adjustments for the application mainly include expanding the state space, optimizing the reward mechanism, enhancing learning rate regulation, and balancing exploration and utilization. The technical roadmap includes basic algorithm analysis, integration of emerging Q-learning technologies, combination of AI assistance methods, and implementation of modular components. From the data in Fig. 6, the path planning of the original algorithm took 8 hours and consumed 250 kilowatt hours. All four Q-learning optimizations improved planning efficiency and accuracy, but "simple parameter adjustment" did not significantly improve. The time was only reduced to 7.5 hours, and the energy consumption was decreased to 245 kilowatt hours. This reflects that algorithm adjustments have not fully adapted to complex processes, and there is a lack of in-depth research on the synergistic benefits of Q-learning and AI. The process optimization effect is shown in Fig. 7.

As shown in Fig. 7, for path planning time, path length, and energy consumption, the model optimized by AI algorithm had the most significant emission reduction effect in scenarios with high relief complexity, followed by scenarios with model structure adjustment. The priority order for optimizing the potential of material waste and unnecessary energy consumption in 3D printing technology was: algorithm structure optimization scenario$\mathrm{>}$AI model training and improvement scenario$\mathrm{>}$parameter fine-tuning scenario. For printing speed, only AI model training could improve efficiency in scenarios, while fine-tuning parameters in scenarios might lead to performance degradation. In order to further verify the application effect of Q-learning algorithm in 3D printing technology of relief technology, a comparative analysis is conducted by introducing traditional genetic algorithm (GA) ^[21] and PSO ^[22]. The results are shown in Table 5.

In Table 5, the performance of Q-learning was significantly better than GA and PSO. In relief process 1, the optimization time of Q-learning was 3.2 hours, with an accuracy of 96.13%, showing significant advantages compared with GA's 5.4 hours and 90.75%, as well as PSO's 6.2 hours and 91.44%. By adaptively learning path planning strategies, Q-learning can converge to the optimal solution faster, reduce the time required for path planning, and improve accuracy. The results of relief process 5 further validated the superiority of Q-learning, with an optimization time of 7.2 hours and an accuracy of 94.23%, which was much better than GA's 11.6 hours and 91.80%, as well as PSO's 13.4 hours and 91.18%. Q-learning enables path planning to achieve high accuracy in a short period of time by continuously iterating and updating strategies. In addition, the optimization results of relief process 9 also showed that Q-learning had a significant improvement in path planning accuracy, with an optimization time of 5.4 hours and an accuracy of 95.12%. It effectively improves the accuracy and efficiency of path planning by continuously adjusting strategies during the learning process. Although the relief process 6 performed average in terms of accuracy improvement, only improving by 0.4%, its time efficiency increased by 14.2 hours, still demonstrating a significant advantage of Q-learning in time optimization. Similarly, the time efficiency improvement of relief process 3 was relatively low, only 1.2 hours, but its accuracy was increased by 1.3%, indicating that Q-learning can also play an important role in small-scale optimization. Therefore, Q-learning performs better than GA and PSO in optimizing the 3D printing path of relief technology, demonstrating its effectiveness in relief technology 3D printing.

4.2 Analysis of the 3D Printing Model Effect of AI Assisted Relief Craft

After analyzing the 3D printing model effect of the relief craft assisted by Q-learning algorithm, the iterative effect of the model is first analyzed. Then the performance of the technology in practical applications is analyzed. The iterative process is shown in Fig. 8.

In Fig. 8, after 40 iterations, the three main evaluation criteria for path planning no longer showed significant changes, indicating that the AI assisted 3D printing path planning for relief craft reached a stable state of optimal solution. From the perspective of path planning performance analysis, the number of iterations significantly affected the speed of reaching a stable optimal solution. Therefore, adjusting the number of iterations appropriately in the Q-learning algorithm iteration craft is beneficial for achieving a stable optimal solution faster. The comparison of path optimization effects in 3D printing is shown in Table 6.

In Table 6, in the 3D printing relief craft, AI assisted path planning utilized Q-learning algorithm to improve the efficiency of printing path and material usage. Although Q-learning and improved GA had similar results in material consumption and printing time, demonstrating the ability to save materials and improve efficiency, Q-learning may be higher in total cost than GA, indicating that it may not necessarily be the best option in cost savings. The limitations of Q-learning in timing optimization may lead to non global optimal solutions. In addition, it shows limitations in considering all variables in the printing process and dealing with multi-objective optimization problems, while the improved GA can more comprehensively balance path, material, and time, providing a more effective and economical relief 3D printing solution. The efficiency of the Q-learning algorithm in 3D printing path planning is shown in Fig. 9.

In Fig. 9, the Q-learning algorithm demonstrated its potential in terms of 3D printing path planning efficiency and material usage. This algorithm performed well in finding the optimal printing path and minimizing material usage. In addition, the Q-learning algorithm exhibited strong convergence ability and high solving efficiency during the iteration process, which is crucial for dealing with complex multi-objective optimization problems. Although there may be room for improvement, the current results have demonstrated the effectiveness of the Q-learning algorithm in path planning for 3D printing models in relief technology.