Numerous energy companies now use generative artificial intelligence (AI) to rapidly retrieve information from hundreds of documents, enable chatbot assistants to increase operational efficiency, and generate synthetic data for use in simulation and decision-making. Out of the various types of generative models, large language models (LLMs) are getting the most attention due to their impressive capabilities for text generation, sentiment analysis, and code generation.

Although LLMs can perform various tasks with natural language, they are generalists, not specialists. Working with generalist language models can be a challenge when processing text data from highly specialized domains that have a language of their own, such the oil and gas industry. Drilling notes, for example, are full of jargon and abbreviations that make the content difficult for nonspecialists to decipher. As oil and gas companies explore how generative AI can help them unlock value from thousands of drilling documents, they will need language models that can understand those documents—including all the jargon and abbreviations.

Amazon Bedrock, a service from Amazon Web Services (AWS) that offers a choice of high-performing foundation models from leading AI companies, has recently launched capabilities to customize foundation models with the user’s own data to build applications that are specific to a domain, organization, and use case. Fine-tuning is a customization approach that increases an LLM’s accuracy by using task-specific labeled data to create a specialized LLM. This blog will illustrate how oil and gas customers can use drilling data to fine-tune an LLM to understand the abbreviations and the terminology of drilling notes and to automatically generate daily summaries. This specialization improves the efficiency of drilling engineers by providing a draft of the report that they must send to oil and gas companies.

Building the solution

The solution uses AWS Glue—a serverless data integration service—and Amazon Athena—a serverless, interactive analytics service—to query drilling reports. It stores the reports using Amazon Simple Storage Service (Amazon S3), which provides object storage built to retrieve virtually any amount of data from anywhere. For its foundation model, the solution uses Amazon Titan Text Express, a customizable foundation model offered through Amazon Bedrock. We will fine-tune Amazon Titan Text Express with hourly comments and daily summaries written by drilling engineers to help the model learn drilling jargon. For fine-tuning, deployment, and inference, we will use the Amazon Bedrock API.

Figure 1. Solution architecture to query drilling reports stored in WITSML format in Amazon S3. An AWS Lambda function converts the XML files into JSON, and an AWS Glue Crawler loads data into the AWS Glue data catalog. Amazon Athena retrieves data using SQL-like statements to fine-tune the model. Both the original model and the customized model are made available to the user for comparison.

Using the Volve dataset

The Norwegian multinational energy company Equinor has made Volve dataset, a set of drilling reports available for research, study, and development purposes. (When using external data, be sure to abide by the license the data is offered under.) The dataset contains 1,759 daily drilling reports—each containing both hourly comments and a daily summary—from the Volve field in the North Sea. Drilling rig supervisors tend to use domain-specific terminology and grammar when describing operations in both the hourly comments and the daily summary. This terminology is standard in the industry, which is why fine-tuning a foundation model using these reports is likely to improve summarization accuracy by enhancing the LLM’s ability to understand jargon and speak like a drilling engineer.

Below is an example daily drilling report from the Volve field in PDF format. Rig operations personnel create daily drilling reports to record the state of the operation, record actions taken earlier that day, and communicate planned activities.

Figure 2. Daily drilling report PDF from the Volve field

In this blog, we will customize Amazon Titan Text Express to specialize in automatically generating a draft of the daily drilling report from the drilling engineers’ annotations. The labeled dataset will consist of the “Summary of activities (24 hours)” section and the “Remark” column in the “Operations” section.

The labeled dataset will be queried from drilling reports from the Volve dataset, which are available in the WITSML format. WITSML is an XML-based hierarchal data format used to store data related to oil and gas operations, based on Energistics standards. To facilitate the querying process, the WITSML files have been transformed into the JSON format and are queried using AWS Glue and Amazon Athena. A few examples of the hourly comments and daily summaries are shown below.

SELECT dr.witsml_drillReports.witsml_drillReport.witsml_nameWell as WellName, 
dr.witsml_drillReports.witsml_drillReport.witsml_dTimStart as start_time, 
dr.witsml_drillReports.witsml_drillReport.witsml_statusInfo.witsml_sum24Hr as  daily_summary_comment,
array_join(transform(dr.witsml_drillReports.witsml_drillReport.witsml_activity, x -> x.witsml_comments),' ') as hourly_comments
FROM "{glue_db}"."{glue_tb}" dr
ORDER BY dr.witsml_drillReports.witsml_drillReport.witsml_nameWell, dr.witsml_drillReports.witsml_drillReport.witsml_dTimStart

Fine-tuning Amazon Titan Text Express using Amazon Bedrock

Amazon Bedrock supports fine-tuning for Meta Llama 2, Cohere Command, and Amazon Titan models through the CreateModelCustomizationJob API method. With this API method, the user specifies the model and the fine-tuning dataset. Optionally, the user can also supply a validation dataset and hyperparameters to customize the training job. For this summarization task, the training data consists of labeled examples, such as the one shown below.

{
    "input": "VERIFIED THAT LPR-N VALVE WASN'T LEAKING.CLOSED LOWER LUBRICATOR VALVE & PERFORMED 10 MINUTE IN-FLOW TEST. CLOSED UPPER LUBRICATOR VALVE. BLED PRESSURE OFF OF SURFACE LINES THROUGH CHOKE MANIFOLD & FLUSHED LINES TO STOCK TANK WITH SEAWATER. WHILE CONTINUING WITH SHUT-IN PRESSURE BUILD-UP TEST, RU ELECTRIC LINE IN PREPARATION FOR FUTURE WIRELINE SAMPLING. CONTINUED PERFORMING SHUT-IN BUILD-UP TEST. CONTINUED PERFORMING SHUT-IN BUILD-UP TEST. OPENED WELL & FLOWED IN PREPARATION OF BOTTOM-HOLE-SAMPLING BHS OPERATIONS. SHUT-IN WELL \u00d8 1645 HRS. RU 4 BHS TOOLSTRING ON ATLAS ELECTRIC LINE. TESTED LUBRICATOR/BOPS TO 200 BAR. RIH WITH BHS TO 3765 M - WELL WAS OPENED-UP AFTER TOOLSTRING WAS BELOW 500 M. SHUT-IN WELL. PERFORM BHS OPERATIONS. POOH WITH BHS / ELECTRIC LINE.\n Summarize the text above",
    "output": "COMPLETED SHUT-IN BUILD-UP TEST. PERFORMED BOTTOM-HOLE-SAMPLING\nOPERATIONS VIA ELECTRIC LINE. STARTED MINI-FRAC OPERATIONS."
}

The CreateModelCustomizationJob method is called with the Amazon S3 path to the training files, as shown below.

create_model_response = bedrock_client.create_model_customization_job(
    jobName="drilling_roport_summarization_job",
    customModelName="drilling_roport_summarization_model",
    roleArn=<arn of IAM role with access to Amazon Bedrock and the S3 training locations>,
    baseModelIdentifier="amazon.titan-tg1-large",
    trainingDataConfig={
        's3Uri': 's3://<s3_bucket_name>/training/drilling_report_training_data.jsonl'
    },
    validationDataConfig={
        'validators': [{
            's3Uri': 's3://<s3_bucket_name>/validation/drilling_report_validation_data.jsonl'
        }]
    },
    outputDataConfig={'s3Uri': 's3://<s3_bucket_name>/finetuning/output'},
    jobTags = [ 
      { 
         "key": "Note",
         "value": "This fine-tuned model summarizes drilling reports"
      }
   ]
)

After the training completes, users can plot the loss and the perplexity curves to observe if the training job converged. Here, we see that after 40 steps, training continues to improve both the training and the validation loss metrics, but perplexity has plateaued for both training and validation data. Often when fine-tuning an LLM, training needs to stop before the loss metrics plateau to preserve the generalizability of the custom model. Otherwise, the resulting model can be so specific to the fine-tuning dataset that it forgets to do other basic tasks—a scenario called catastrophic forgetting.

Figure 3. Loss and perplexity curves for the example data

Evaluating the custom model

The custom model’s ability to generate daily summaries that resemble those written by drilling engineers will be evaluated both quantitatively and qualitatively. For a quantitative evaluation, we will use the Bilingual Evaluation Understudy (BLEU) metric—a numerical metric describing text similarity that is commonly used to measure the quality of translated text—to measure the similarity between summaries prepared by a language model and those created by drilling engineers. The BLEU score ranges from 0 to 1, with higher values indicating a better match between the texts. In other words, the closer the value is to 1, the better the generated text. On the qualitative side, we will select a few examples to observe if the language models use proper industry terminology and acronyms and if they can select the most relevant activities to include in the daily summary.

The table below shows several completions from both the foundation model and the fine-tuned model and how they compare to the summaries generated by a rig supervisor. The fine-tuned completion tends to have a better match to the human summary in terms of content (that is, the fine-tuned model has better judgment in selecting relevant activities to be included in the summary), length, and style. The BLEU score also indicates that the custom model has a much closer resemblance to the summaries prepared by drilling engineers than the original model, as shown by the score increasing by several orders of magnitude in some cases.

Cleaning up

If you deploy a custom model using Amazon Bedrock, remember afterward to navigate to the Custom Models tab on the Amazon Bedrock AWS Console page to delete the model and to prevent incurring cost for models that you are not using.

Conclusion

Orchestrating, parallelizing, and managing fine-tuning jobs to customize foundation models for specific tasks used to be difficult and expensive. Amazon Bedrock simplifies this process so that anyone can build, deploy, and maintain sophisticated fine-tuned models that understand and respond in a domain-specific verbiage. AWS customers can quickly fine-tune a foundation model without dedicated data scientists or extensive expertise.

Generative AI has the potential to improve efficiency by automating time-consuming tasks even in domains that require deep knowledge of industry-specific nomenclature and acronyms. Having a custom model that provides drilling engineers with a draft of daily activities has the potential to save hours of work every week. Model customization can also help energy and utilities customers in other applications that involve the generation of highly technical content, as is the case of geological analyses, maintenance reports, and shift handover reports.