To address the systematic errors inevitably introduced during the physical fabrication and assembly of multilayer diffractive deep neural networks （D

2

NN）， this paper aims to validate and implement a physically aware adaptive training strategy. By incorporating programmable layers into the diffractive architecture， an optoelectronic closed-loop feedback system is constructed to enable online correction and compensation for wavefront aberrations in the actual physical optical path. The objective of this study is to quantitatively evaluate the effectivenes

s of this strategy in mitigating physical environmental non-idealities， ensuring high reliability and inference fidelity during practical deployment.The proposed method constructs an optoelectronic closed-loop feedback framework with a hybrid “fixed layer + programmable layer” architecture based on physically aware adaptive training. During the physical forward propagation phase， detectors conjugate to the modulation surface are introduced within the diffraction layer to capture the intensity distribution of actual distortions. During backward updating， a pre-trained U-Net is employed as a digital phase-recovery module to reconstruct the complex amplitude， which is then used to estimate physical gradients and iteratively update the amplitude modulation parameters of the digital micromirror device （DMD） layer， thereby enabling reconstruction of the degraded wavefront. To comprehensively validate this approach， this paper employs a combined simulation and experimental methodology： First， an end-to-end simulation environment is established using the angular spectrum propagation model with parameters matched to the physical system. Subsequently， a corresponding physical optoelectronic experimental platform is constructed， integrating a high-precision 3D-printed fixed phase layer， a DMD serving as a programmable diffractive layer， a complementary metal-oxide-semiconductor （CMOS） detector， and a computer control unit to implement a complete closed-loop adaptive compensation strategy. Experiments are performed on the Modified National Institute of Standards and Technology （MNIST） binary classification task and the Fashion-MNIST four-class classification task at a wavelength of 638 nm to systematically evaluate the dynamic compensation capability and fault tolerance of the proposed architecture under mechanical displacement errors.The results indicate that， under ideal error-free conditions， simulation analysis shows that the system can achieve a theoretical classification accuracy of over 95% in the MNIST binary classif

ication task. When a two-dimensional displacement error of 5 pixels （9 μm） is introduced in both the x and y directions across all fixed phase layers， the classification performance of the uncompensated system degrades significantly. By adopting an adaptive training strategy， the model converged rapidly within 10 iteration cycles， restoring the accuracy to a level close to the ideal. Furthermore， under extreme displacement errors of 8 pixels， this strategy successfully improved the initial accuracy from below 20% to approximately 90% while maintaining system stability. In the MNIST physical experiment， faced with real-world alignment deviations and manufacturing errors， the system improved the initial recognition accuracy from 81% to 92% by iteratively optimizing the DMD modulation parameters. Even when 3-pixel and 6-pixel two-dimensional displacement disturbances were artificially introduced， the network maintained stable classification performance， effectively mitigating the negative effects. To further validate the universality of this strategy under complex decision boundaries， this study conducted additional validation on the Fashion-MNIST four-class classification task. Compared to simple MNIST digits， clothing images possess more complex spatial texture features， placing higher demands on the accuracy of wavefront reconstruction. Experimental results show that the adaptive compensation strategy successfully increased the experimental inference accuracy for this task from an initial 76% to 88%. The spot， which had previously diverged due to error accumulation， refocused within the target detection area following adaptive optimization. This successful validation across datasets provides physical evidence of the architecture’s robust adaptability when handling high-dimensional feature extraction tasks.In summary， the physically aware adaptive training strategy effectively alleviates the performance degradation of multilayer D2NNs caused by physical misalignment in practical applications. By introducing a DMD-

based compensation mechanism， the system effectively neutralizes multi-scale macro-level misalignment errors. Consistency between simulations and experiments fully demonstrates that this optoelectronic closed-loop feedback architecture significantly narrows the performance gap between digital models and physical entities. This approach does not rely on stringent hardware alignment accuracy or small perturbation assumptions， providing an efficient and flexible technical pathway for robust computation of diffractive neural networks in non-ideal physical environments.