Text-Guided Well Log-Constrained Realistic Subsurface Model Generation via Stable Diffusion

Building subsurface models requires the incorporation of geological knowledge that often comes in the form of text. Such knowledge is typically first converted into mathematical formulation and then used as part of an inverse problem as a penalty or constraint. Stable diffusion models can be used for conditional image generation thus digitizing the geological information into spatially varying priors that can later be used in inversion. Here we show how such priors can be generated and conditioned using well logs. In particular, we create a new small dataset where we extract patches from well-known public seismic models. The dataset of less than 50 samples extracted from public geological models and annotated using a large language model with vision capability is sufficient to fine-tune an open-source stable- diffusion pipeline. This approach allows a new way of working with geological descriptions or any other information with minimal effort.

1. Text-Guided Well Log-Constrained High-Fidelity Subsurface Model Generation via Stable Diffusion Oleg Ovcharenko 1 , Vladimir Kazei 2 , Weichang Li 2 , Issam Said 1 1 – NVIDIA, Dubai, UAE 2 – Aramco Americas, Houston Research Center

2. Generation of random subsurface models from text description ”Anticlinal structure with potential reservoir below a salt dome"

3. Why do we need synthetic data? Train DL * models Create benchmarks

4. Problem Solution Latent diffusion Low-Rank Adaptation Method Data generation Training Results Conclusions

5. Problem Subsurface models Synthetic data Inverted data Well-logs Metadata Text Solver Training Text description of the desired model distribution is missing out

6. Solution Need to merge knowledge from different modalities Text  Image Leverage diffusion model for conditional image generation Example generated with OpenAI DALL-E Subsurface models Text

7. Getting Lost in the Noise Forward Diffusion t = 0 t = 1 t = 2 t = 3 t = 4 t = 5 t = 6 t = 7 t = 8 t = 9 t = 10   Generate noise from a standard normal distribution Multiply result by   Step 1: Multiply image at previous step by Add it to the result from step 1   Step 2: noise = torch.randn_like ( x_t ) x_t = torch.sqrt (1 - B[t]) * x_t + torch.sqrt (B[t]) * noise Code:

8. Getting Lost in the Noise Reverse Diffusion t = 0 t = 1 t = 2 t = 3 t = 4 t = 5 t = 6 t = 7 t = 8 t = 9 t = 10   The noise added at time This is what the neural network will estimate                                              

9. Model Training Training Data Noisy image at time   Noise added at time     “Skip ahead” function      

10. Down0 Copy Down1 Copy It’s About Time Feature Map (Down2) Feature Map (Down1) Latent Vector Down2 Copy Feature Map (Up0) Feature Map (Up1) Feature Map (Up2) Feature Map (Down0) Adding Time Time Time Embed 2 Time Embed 1

11. LoRA https://arxiv.org/pdf/2106.09685

12. Image b y Aayush Agrawal https://towardsdatascience.com/stable-diffusion-using-hugging-face-501d8dbdd8

13. Make text and image embeddings CLIP Contrastive Language-Image Pre-Training VAE Variational Autoencoder Image b y Aayush Agrawal https://towardsdatascience.com/stable-diffusion-using-hugging-face-501d8dbdd8

14. Dataset creation: fetch public subsurface models 2 D 3D https://wiki.seg.org/wiki/ SEG Salt model Overthrust 3D Marmousi II BP 2004 Sigsbee

15. Dataset creation: sample crops from base models Keep aspect ratio but rescale to min side of 512 Total 45 images after QC

16. Dataset creation: prompt for image annotation "Overhanging salt dome with flank collapse and potential hydrocarbon trap at crest." "Thrust faulting with folded sedimentary layers and potential hydrocarbon trap in anticlinal structure."

17. Dataset creation: prompt for image annotation Na ï ve image descriptions are far from ideal. Context-aware annotation is the way

18. Training takes 5 min on 1 GPU 11 121 231 451 1551

19. Condition generation by well-log and text

20. Results "Anticline fold layered structure with potential hydrocarbon traps" (t/l) "Flat sedimentary layers with angular unconformities" (t/r) "Salt dome intrusion with potential hydrocarbon traps" (b/l) "Complex layered geological environment” (b/r) Text and Log conditioned Text conditioned only

21. How to accelerate? https://github.com/NVIDIA/TensorRT/tree/release/8.6/demo/Diffusion

22. https://github.com/NVIDIA/TensorRT/blob/release/9.3/demo/experimental/HuggingFace-Diffusers/TensorRT-diffusers-txt2img.ipynb

23. Conclusion Stable diffusion allows to incorporate textual geological descriptions as a source of information into numerical subsurface model building Iterative nature of stable diffusion is a computational bottleneck Performance-efficient fine tuning (PEFT) methods such as LoRA are suitable to introduce domain knowledge even with limited data sets Diffusion hyperparameters affects the fidelity of generated data