Valued for its nondestructive， non-contact operation and high precision， spectroscopic ellipsometry has established itself as the standard characterization method for measuring the thickness and optical constants of nanoscale thin films and is routinely employed in the quality inspection of coatings. The conventional spectroscopic ellipsometry demodulation approach is model-based. It involves constructing a pre-defined optical model for the thin film， complete with a dielectric function model， and then iteratively refining the model parameters via nonlinear regression analysis to extract the properties of the film. However， the complexity of this modeling and the iterative solving process result in low computational efficiency， making it difficult to meet the demands for rapid inspection of coated samples. Neural networks can establish a direct mapping from ellipsometri

c spectra to thin-film properties， thereby enabling rapid data demodulation.However， the application of neural networks in spectroscopic ellipsometry data demodulation is constrained by the challenges of obtaining sufficient training data and ensuring its high quality.To address those challenges， this paper proposes a neural network-based demodulation method for spectroscopic ellipsometry data of non-absorbing thin films， incorporating the Cauchy dispersion model as a physical constraint. In this method， a dataset of ellipsometric spectra was first generated based on the Cauchy dispersion model and the theory of polarized light interference in thin films. Subsequently， a neural network was trained using this dataset， with the film thickness and the parameters of the Cauchy model designated as its output targets. The dataset generation procedure was conducted as follows. First， the value ranges for the Cauchy model parameters （

A

，

B

， and

C

） were established based on the material's attainable refractive index range and its spectral variation. Within these ranges， parameters A and B were sampled at equal intervals， and C was randomly sampled. All possible combinations of （

A

，

B

and

C

） were then traversed to generate the corresponding refractive index data via the Cauchy dispersion model. Subsequently， the film thickness

d

was sampled at equal intervals within its predefined range. Each sampled thickness value was paired with every generated refractive index dataset. Following this， the ellipsometric spectra for each （

d

，

n

） pair were simulated according to the principles of polarized light interference in thin films. Finally， experimental noise was added to these ideal spectra to produce the final ellipsometric spectra dataset for network training. In this work， the ResNet-18 architecture was selected as the backbone network. It was specifically configured to be one-di

mensional to accommodate the sequential nature of spectroscopic ellipsometry data. The two measured ellipsometric parameters are used as the input of network. The outputs of the network consist of the film thickness d and the three Cauchy model parameters （

A

，

B

and

C

）. The wavelength-dependent refractive index is then reconstructed from these parameters using the Cauchy dispersion model. This design， as opposed to having the network directly output the refractive index at multiple discrete wavelengths， effectively reduces the complexity of model， and improves its learning efficiency.Experimental validation on four distinct materials （MgF

2

， SiO

2

， Si

3

N

4

， and TiO

2

） demonstrates that the proposed network achieves an average thickness demodulation error of less than 1.5 nm， an average refractive index error below 0.02， and an average computation time of approximately 6 ms. It significantly outperforms the conventional Levenberg-Marquardt （LM） algorithm in both accuracy and efficiency， demonstrating substantial advantages in both performance and stability for rapid characterization. The proposed demodulation framework also demonstrates the advantage of enabling targeted data augmentation. Specifically， tailored data augmentation strategies were implemented for two distinct material categories： low-dispersion， low-refractive-index materials and high-dispersion， high-refractive-index materials. This process yielded two specialized network models. Compared to the original general-purpose model， these specialized models demonstrate superior demodulation performance on their respective target material types. Finally， an analysis was performed to investigate the impact of material absorption on the demodulation accuracy and to determine the corresponding tolerance threshold. The results indicate that the absorption characteristic of material generally degrades the demodulation precision. Nevertheless， the proposed method

maintains satisfactory accuracy when the extinction coefficient k is below 0.05， establishing a practical applicability boundary for absorption-free or weakly absorbing thin films.In conclusion， the proposed neural network framework addresses the two challenges of training data acquisition and fast demodulation in spectroscopic ellipsometry， offering a novel and efficient pathway for thin-film characterization.