Drill hole constraint geological section generation method based on diffusion generation type model
By training a mask-constrained diffusion generative model based on a diffusion generative model, the problems of subjectivity and insufficient representation of complex structures in borehole-constrained geological profile prediction are solved, generating clear geological profile images that meet engineering requirements and improving the stability and diversity of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTHWEST JIAOTONG UNIV
- Filing Date
- 2026-01-20
- Publication Date
- 2026-04-28
AI Technical Summary
Existing technologies for predicting borehole-constrained geological profiles suffer from strong subjectivity, difficulty in representing complex structures, and insufficient characterization of multiple solutions and uncertainties. Furthermore, generative models involve trade-offs between training stability, diversity, and resolution, making it difficult to generate clear and smooth stratigraphic color bands that meet actual engineering needs.
A diffusion-based generative model approach is adopted. By constructing a binary mask image and an occluded input image, a mask-constrained diffusion generative model is trained. The Stable-Diffusion-2inpainting network structure is used in combination with a variational autoencoder and a U-Net network to generate a complete geological profile image that meets borehole constraints.
It enables the automatic generation of clear and smooth geological profile images with stratigraphic color bands that meet engineering requirements under borehole constraints, improving the training stability of the model and the diversity of generated results, and is able to systematically characterize the uncertainty of underground structures.
Smart Images
Figure CN121937571A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for generating borehole-constrained geological profiles based on a diffusion-generative model, belonging to the field of geological engineering and information processing technology. Background Technology
[0002] Geological profiles are crucial outputs reflecting underground stratigraphic structure, lithological assemblage, and spatial continuity, playing a fundamental role in fields such as engineering site stability assessment, foundation pit and underground engineering design, tunnel alignment, and oil, gas, and mineral exploration. Traditional geological profile construction primarily relies on the following methods: On the one hand, engineering geologists typically rely on borehole columnar sections, combined with surface geological survey results and regional geological background, to manually draw stratigraphic boundaries along the profile lines, thus forming a schematic diagram of stratigraphic zonation. This method is highly dependent on the experience of engineers; different personnel may provide significantly different profile plans for the same borehole combination, exhibiting strong subjectivity and making it difficult to systematically depict the uncertainties of underground structures.
[0003] On the other hand, numerical methods such as random field theory, kriging interpolation, and multi-point statistics (MPS) have also been used for geological modeling under borehole constraints. These methods typically treat stratigraphic properties as random fields, using covariance functions, training images, or conditional simulation techniques to perform spatial interpolation and stochastic simulation between boreholes. While these methods can express multi-point spatial relationships to some extent using training images, their ability to represent complex stratigraphic geometries (such as pinch-outs, lenticular bodies, meandering channels, and multi-stage superimposed sediments) is limited. Furthermore, they mostly operate in scalar or categorical field spaces, making it difficult to directly output complete profile images with clear color bands, boundary morphology, and visualization effects.
[0004] In recent years, generative models such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) have been attempted for generating and augmenting geological or training images. However, there is a trade-off between training stability, diversity, and resolution, making it difficult to directly transfer the model structure and training process to the specific geological scenario of "borehole-constrained 2D profile prediction." Especially in practical engineering, it is required that the generated results strictly meet the hard constraints of boreholes, while also hoping that the generated geological profiles have clear and smooth stratigraphic color bands, rather than the complex textures and lighting effects found in natural images. Current technologies lack specialized methods to address this requirement. Summary of the Invention
[0005] The purpose of this invention is to overcome the problems of strong subjectivity, difficulty in expressing complex structures, and insufficient characterization of multiple solutions and uncertainties in existing methods for predicting borehole-constrained geological profiles, and to propose a method for generating borehole-constrained geological profiles based on a diffusion generative model.
[0006] The technical solution provided by this invention to solve the above-mentioned technical problems is: a method for generating borehole-constrained geological profiles based on a diffusion-generative model, comprising the following steps: Step S1: Obtain the original geological profile image data and preprocess the original geological profile image data; Step S2: Construct a binary mask image on the original geological profile image and generate an occluded input image, then construct a sample training set; Step S3: Take the binary mask image and the occluded input image as input, and the original geological profile image as output, and train the mask-constrained diffusion generation model in the latent space to obtain a diffusion generation model that can generate two-dimensional geological profiles under the constraints of borehole information and mask conditions. Step S4: Project the boreholes in the geological profile to be predicted along the profile line to be studied and rasterize them. Fill the different lithological sections with color and write them into the corresponding raster columns to obtain the occluded input image with stratigraphic information only at the borehole location. Generate a corresponding binary mask image based on whether the borehole exists. Step S5: Encode the occluded input image and binary mask image obtained in step S4 into a latent space tensor, and input it together with the randomly initialized noisy latent space tensor into the mask-constrained diffusion generation model trained in step S3, and perform a back-diffusion sampling process to obtain a complete geological profile image that satisfies borehole constraints.
[0007] A further technical solution is that the specific process of preprocessing in step S1 is as follows: unify the original geological profile image data to preset spatial coordinates and image resolution, and represent different lithological or stratigraphic units using discrete color or label encoding.
[0008] A further technical solution is that, in step S2, several grid columns are selected in the horizontal direction of the complete geological profile image as visible borehole zones, and a binary mask image is constructed. In the binary mask image, the pixels corresponding to the visible borehole zones are assigned a first value, and the remaining pixels are assigned a second value. The regions in the complete geological profile image whose mask values are the second value are replaced with neutral fill values to obtain an occluded input image. Thus, a sample training set consisting of the complete geological profile image, the mask image, and the occluded input image is obtained.
[0009] A further technical solution is that the neutral fill value in step S2 is a single grayscale value or a single color value.
[0010] A further technical solution is that, in step S2, the rule for selecting several grid columns in the horizontal direction of the cross-section as the visible borehole zone includes at least one of the following: (1) Arrange virtual borehole rows at fixed horizontal intervals; (2) The borehole rows are arranged according to the horizontal position and depth of the actual engineering boreholes; (3) Randomly sample the positions of the borehole columns in the horizontal direction and randomly extract the borehole depth to generate multiple mask patterns to enhance the adaptability of the training mask constraint diffusion generation model to different borehole combinations.
[0011] A further technical solution is that the mask-constrained diffusion generation model is constructed based on the Stable-Diffusion-2inpainting network structure, in which the parameters of the variational autoencoder and the text encoder are kept frozen during training, and only the parameters of the U-Net network are fine-tuned.
[0012] A further technical solution is that the training process in step S3 is as follows: Step S31: Initially generate a mask-constrained diffusion generation model; Step S32: In each training step, the occluded input image and the binary mask image are fed into the model, and the noise prediction error in the latent space is backpropagated and the parameters are updated so that the model gradually approximates the complete profile image under the given mask constraints. Step S33: After every few training rounds, generate a prediction profile on the validation set and calculate the structural similarity (SSIM), peak signal-to-noise ratio (PSNR), or pixel classification accuracy to obtain the prediction accuracy curve that changes with the number of training rounds. Step S34: During the training process, the model weight parameters are periodically saved to form multiple checkpoints, which facilitates interruption of training and selection of the best model. When the evaluation index on the validation set tends to converge or reaches the preset threshold, training is stopped, and the checkpoint with the best index is selected as the final mask constraint diffusion generation model.
[0013] A further technical solution is that the specific process of the training step in step S32 is as follows: Step S321: Encode the high-resolution image using a variational autoencoder to obtain the latent variables corresponding to each complete profile image; Step S322: Perform diffusion modeling in the low-dimensional latent space to obtain noisy latent variables; Step S323: At each time step, the noisy latent variable time step encoding and image inpainting conditions are used as inputs to U-Net, and the network output is a prediction of the noise to obtain the denoised latent variables. Step S324: Input the denoised latent variables into the decoder to restore them to the pixel space and obtain the predicted complete geological profile image.
[0014] A further technical solution is that the specific process of step S4 is as follows: First, obtain borehole columnar section data and project the borehole spatial coordinates onto the selected profile line, discretizing them into a two-dimensional grid; according to the borehole depth and stratigraphic division, fill different lithological sections into the corresponding grid columns to obtain a masked input image with stratigraphic information only at the borehole location; at the same time, construct a binary mask image based on whether borehole information exists in the profile grid.
[0015] The beneficial effects of this invention are as follows: This invention transforms historical or simulated geological profiles into training samples of "virtual borehole constraints - complete profiles" and trains a mask constraint diffusion generation model in the latent space, so that in practical applications, it can automatically generate complete geological profile maps that meet borehole constraints based on given borehole columnar data, and obtain a set of profile samples through multiple samplings when needed, which can be used to construct lithological probability profile maps and evaluate the uncertainty of geological structures. Attached Figure Description
[0016] Figure 1 This is the original geological profile. Figure 2 The training images are after image preprocessing and size normalization; Figure 3 For masked images; Figure 4 Drilling diagram for implementation case; Figure 5 The prediction accuracy curve varies with the number of training rounds; Figure 6 Geological profile map showing the prediction results of this method; Figure 7 This is the overall flowchart. Detailed Implementation
[0017] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] The present invention provides a method for generating borehole-constrained geological profiles based on a diffusion-generative model, comprising the following steps: Step S1: Obtain the original geological profile image data and preprocess the original geological profile image data; The specific process is as follows: unify the original geological profile image data to a preset spatial coordinate and image resolution (e.g., 512×512 pixels), and use discrete color or label encoding to represent different lithological or stratigraphic units; Step S2: Construct a binary mask image on the original geological profile image and generate an occluded input image, then construct a sample training set; The specific process is as follows: In the complete geological profile image, according to the preset borehole layout rules, several grid columns are selected in the horizontal direction to construct a binary mask image. In the binary mask image, the pixels corresponding to the visible borehole zone are assigned a first value, and the remaining pixels are assigned a second value (for example, the position with a mask of 1 represents the visible borehole zone, and the position with a mask of 0 represents the process domain to be completed). The area in the complete geological profile image with the mask value of the second value (i.e., the area with a mask of 0) is filled and replaced with neutral fill (single gray value or single color value) to obtain the occluded input image. Thus, a sample training set composed of the complete geological profile image, the mask image, and the occluded input image is obtained. The preset drilling layout rules include at least one of the following: (1) Arrange virtual borehole rows at fixed horizontal intervals; (2) The borehole rows are arranged according to the horizontal position and depth of the actual engineering boreholes; (3) Randomly sample the positions of the borehole columns in the horizontal direction and randomly extract the borehole depth to generate multiple mask patterns to enhance the adaptability of the training mask constraint diffusion generation model to different borehole combinations; Step S3: Take the binary mask image and the occluded input image as input, and the original geological profile image as output, and train the mask-constrained diffusion generation model in the latent space to obtain a trained mask-constrained diffusion generation model that can generate two-dimensional geological profiles under borehole and mask conditions. The mask-constrained diffusion generative model is built on the Stable-Diffusion-2 inpainting network structure, in which the parameters of the variational autoencoder and text encoder are kept frozen during training, and only the parameters of the U-Net network are fine-tuned.
[0019] The specific training process is as follows: Step S31: Initially generate a mask-constrained diffusion generation model; Step S32: In each training step, the occluded input image and the binary mask image are fed into the model, and the noise prediction error in the latent space is backpropagated and the parameters are updated so that the model gradually approximates the complete profile image under the given mask constraints. Step S321: Encode the high-resolution image using a variational autoencoder to obtain the latent variables corresponding to each complete profile image. ; The encoder is denoted as The decoder is Then each complete cross-sectional image The corresponding latent variables are: ,
[0020] Step S322: Perform diffusion modeling in the low-dimensional latent space to obtain noisy latent variables. ; Latent variables given a complete profile The forward diffusion process gradually adds noise through a fixed noise schedule, and its expression is:
[0021]
[0022] in This is the cumulative factor calculated based on a preset noise schedule; Step S323: Then, at each time step t, use noisy latent variables. The time-step encoding and image inpainting conditions are used as inputs to U-Net, causing the network output to predict noise, thus obtaining the denoised latent variables. ; Step S324: Denoise the latent variables The input is decoded to restore the pixel space, resulting in a predicted complete geological profile image. Step S33: After every few training rounds, generate a prediction profile on the validation set and calculate the structural similarity (SSIM), peak signal-to-noise ratio (PSNR), or pixel classification accuracy to obtain the prediction accuracy curve that changes with the number of training rounds. Step S34: During the training process, the model weight parameters are periodically saved to form multiple checkpoints, which facilitates interruption of training and selection of the best model. When the evaluation index on the validation set tends to converge or reaches the preset threshold, training is stopped, and the checkpoint with the best index is selected as the final mask constraint diffusion generation model.
[0023] The loss function is:
[0024] Where cond represents the conditional input consisting of masked latent variables and some visible latent variables; Step S4: Project the boreholes in the geological profile to be predicted along the profile line to be studied and rasterize them. Fill the different lithological sections with color and write them into the corresponding raster columns to obtain the occluded input image with stratigraphic information only at the borehole location. Generate a corresponding binary mask image based on whether the borehole exists. First, borehole columnar section data is acquired, and the spatial coordinates of the boreholes are projected onto the selected profile line and discretized into a two-dimensional raster grid. Based on the borehole depth and stratigraphic division, different lithological sections are filled with color and written into the corresponding raster columns to obtain a masked input image with stratigraphic information only at the borehole location. At the same time, a binary mask image is constructed based on whether borehole information exists in the profile grid.
[0025] Step S5: Encode the occluded input image and binary mask image obtained in step S4 into a latent space tensor, and input it together with the randomly initialized noisy latent space tensor into the mask-constrained diffusion generation model trained in step S3, and perform the reverse diffusion sampling process to obtain a complete geological profile image that satisfies the borehole constraints. Conditional diffusion generates a complete geological profile: during the generation stage, initial latent variables are sampled from an isotropic Gaussian distribution. Combining a pre-trained mask-constrained diffusion generation model, and following the backsampling formula in DDPM / LatentDiffusion, from time step... Iterate step by step to 0 to complete the denoising process; each time step update uses the same inverse update operator as the original mask-constrained diffusion generation model, finally obtaining the denoised latent variables. Finally, Input Decoder By restoring the image to pixel space, a complete geological profile image that satisfies the borehole constraints is obtained: .
[0026] The above description is not intended to limit the present invention in any way. Although the present invention has been disclosed through the above embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall fall within the scope of the present invention.
Claims
1. A method for generating borehole-constrained geological profiles based on a diffusion-generative model, characterized in that, Includes the following steps: Step S1: Obtain the original geological profile image data and preprocess the original geological profile image data; Step S2: Construct a binary mask image on the original geological profile image and generate an occluded input image, then construct a sample training set; Step S3: Take the binary mask image and the occluded input image as input, and the original geological profile image as output, and train the diffusion generation model constrained by the spatial mask in the latent space to obtain a diffusion generation model that can generate two-dimensional geological profiles under the constraints of borehole information and mask conditions. Step S4: Project the boreholes in the geological profile to be predicted along the profile line to be studied and rasterize them. Fill the different lithological sections with color and write them into the corresponding raster columns to obtain the occluded input image with stratigraphic information only at the borehole location. Generate a corresponding binary mask image based on whether the borehole exists. Step S5: Encode the occluded input image and binary mask image obtained in step S4 into a latent space tensor, and input it together with the randomly initialized noisy latent space tensor into the diffusion generation model with mask constraints trained in step S3, and perform a back-diffusion sampling process to obtain a complete geological profile image that satisfies borehole constraints.
2. The method for generating borehole-constrained geological profiles based on a diffusion-generative model according to claim 1, characterized in that, The specific preprocessing process in step S1 is as follows: unify the original geological profile image data to preset spatial coordinates and image resolution, and represent different lithological or stratigraphic units using discrete color or label encoding.
3. The method for generating borehole-constrained geological profiles based on a diffusion-generative model according to claim 1, characterized in that, In step S2, several grid columns are selected in the horizontal direction of the complete geological profile image as visible borehole zones, and a binary mask image is constructed. In the binary mask image, the pixels corresponding to the visible borehole zones are assigned a first value, and the remaining pixels are assigned a second value. The regions in the complete geological profile image whose mask values are the second value are replaced with neutral fill values to obtain an occluded input image. Thus, a sample training set consisting of the complete geological profile image, the mask image, and the occluded input image is obtained.
4. The method for generating borehole-constrained geological profiles based on a diffusion-generative model according to claim 3, characterized in that, The neutral fill value in step S2 is either a single grayscale value or a single color value.
5. The method for generating borehole-constrained geological profiles based on a diffusion-generative model according to claim 3, characterized in that, In step S2, the rule for selecting several grid columns in the horizontal direction of the profile as the visible borehole zone is as follows: including at least one of the following: (1) Arrange virtual borehole rows at fixed horizontal intervals; (2) The borehole rows are arranged according to the horizontal position and depth of the actual engineering boreholes; (3) Randomly sample the positions of the borehole columns in the horizontal direction and randomly extract the borehole depth to generate multiple mask patterns to enhance the adaptability of the training mask constraint diffusion generation model to different borehole combinations.
6. The method for generating borehole-constrained geological profiles based on a diffusion-generative model according to claim 1, characterized in that, The mask-constrained diffusion generation model is built on the Stable-Diffusion-2 inpainting network structure, in which the parameters of the variational autoencoder and text encoder are kept frozen during training, and only the parameters of the U-Net network are fine-tuned.
7. The method for generating borehole-constrained geological profiles based on a diffusion-generative model according to claim 6, characterized in that, The training process in step S3 is as follows: Step S31: Initially generate a mask-constrained diffusion generation model; Step S32: In each training step, the occluded input image and the binary mask image are fed into the model, and the noise prediction error in the latent space is backpropagated and the parameters are updated so that the model gradually approximates the complete profile image under the given mask constraints. Step S33: After every few training rounds, generate a prediction profile on the validation set and calculate the structural similarity (SSIM), peak signal-to-noise ratio (PSNR), or pixel classification accuracy to obtain the prediction accuracy curve that changes with the number of training rounds. Step S34: During the training process, the model weight parameters are periodically saved to form multiple checkpoints, which facilitates interruption of training and selection of the best model. When the evaluation index on the validation set tends to converge or reaches the preset threshold, training is stopped, and the checkpoint with the best index is selected as the final mask constraint diffusion generation model.
8. The method for generating borehole-constrained geological profiles based on a diffusion-generative model according to claim 7, characterized in that, The specific process of the training step in step S32 is as follows: Step S321: Encode the high-resolution image using a variational autoencoder to obtain the latent variables corresponding to each complete profile image; Step S322: Perform diffusion modeling in the low-dimensional latent space to obtain noisy latent variables; Step S323: At each time step, use the noisy latent variable time step encoding and the image inpainting conditions as input to U-Net, and make the network output the prediction of noise to obtain the denoised latent variable. Step S324: Input the denoised latent variables into the decoder to restore them to the pixel space and obtain the predicted complete geological profile image.
9. The method for generating borehole-constrained geological profiles based on a diffusion-generative model according to claim 1, characterized in that, The specific process of step S4 is as follows: First, obtain borehole columnar section data and project the borehole spatial coordinates onto the selected profile line, discretizing them into a two-dimensional grid. According to the borehole depth and stratigraphic division, fill different lithological sections into the corresponding grid columns to obtain a masked input image with stratigraphic information only at the borehole location. At the same time, construct a binary mask image based on whether borehole information exists in the profile grid.