Digital twins and Large Language Models (LLMs) arose in two separate disciplines. Digital twins were developed to streamline complex workflows in agile project lifecycle management on top of structured data. LLMs emerged from AI research for translating unstructured data into different formats and contexts, driving the recent enthusiasm for generative AI.
Most discussions about digital twins typically focus on concrete things like buildings, bodies, or physical products. But they can equally be applied to make sense of more abstract things like financial portfolios, customers, or business processes.
Mamdouh Refaat, Chief Data Scientist and Senior VP of Product Management at Altair believes these two approaches can complement each other’s limitations in the banking and financial services industry (BFSI). LLMs can translate unstructured data into structured data sets more suitable for complex modeling and collaboration in digital twins. Meanwhile, digital twins can provide a stronger foundation for LLMs to reduce hallucinations and improve accuracy, he explains:
“We anticipate the first wave of the integration of LLMs with digital twins to be through providing a natural language interface to navigate and explore digital twins. The next phase will be using LLMs to generate data related to the content of digital twins that can be used in the modeling and control of digital twins.”
For example, if a bank has a digital twin of a credit card portfolio to monitor its default rate and simulate the impact of changes in the economy, an LLM can be used to generate the data to investigate these changes. The user would interact with the digital twin in a conversation style to run the simulation using the generated data and the LLM to generate a report explaining the results. The role of the human user will be to focus on the orchestration of the requests and on validating and interpreting the results to take the appropriate actions.
New processes required
Although the combination shows tremendous promise, it’s not currently realistic to bolt ChatGPT onto a digital twin platform and hope for the best. Refaat argues that enterprises will also need to undergo some structural and organizational changes to see significant benefit.
For example, currently, BFSI organizations implement digital twins for credit risk scoring in a semi-automated fashion. However, behind the scenes is an army of data scientists developing the ML models, validating, testing, and monitoring their performance. A significant portion of the content is generated by business analysts to support the marketing and management of the credit risk operation.
When LLMs are integrated into this process, more content and data will be generated automatically. However, there will also be an increased demand for human curation of this content and data to guarantee accuracy and relevance. As a result, data scientists will spend more time testing and validating the new data extracted from the content generated by LLMs. Refaat says:
This shift in roles will require changes in the organizational structure of different functions and the flow of data and decisions. There will also be new roles and responsibilities in the organization to control the sources of data to be used to generate content. The governance of the use and deployment of data and content generated by LLMs will also be another area where new rules and operating procedures will need to be set. For example, LLMs can now generate computer code to perform different tasks. The quality assurance of this code will need to be managed using new procedures. One may even ask, ‘Could LLMs then be used to generate these changes?’
From concrete to abstract
Digital twin tools provide a framework for generating digital representations of real dynamic systems. In concrete domains like construction or product engineering, the model examines the problem’s physics or uses a data-driven model derived with machine learning. Digital twins can also receive data from the real system and use it to update the model or compare it to the output of the model to make decisions on how to control or influence the real model.
In domains like banking, finance, and insurance, the models are a bit more abstract than physical things. For example, banks regularly monitor the health of their credit card portfolios, often measured by the overall default rate. The default rate depends on the profile of the account holders and the general economic indicators such as interest rates, inflation, unemployment rates, global events, such as the COVID-19 pandemic, and so on.
Using this data, the bank can develop a machine learning model that simulates the effect of these variables on the default rate of their credit card portfolios and run a what-if analysis to assist in setting up new portfolios and customer management strategies. This model is essentially a digital twin of the credit card portfolio.
This is where LLMs come in. They use text to make predictions based on text coming in. These are generative machine learning models that generate data in the form of text as their output. For example, the GPT model underpinning ChatGPT is trained using text from the internet to understand text input and generate a predicted text for different purposes. This helps it answer questions about a subject using a specific scope of text data, such as newspapers or magazine articles, it finds the most likely predicted answer from those content sources. It can also translate a large corpus of financial documents into the appropriate format for more structured analysis and modeling.
Old wine in new models
Refaat argues that digital twins are not new to BFSI. The banking industry has used similar principles, albeit with different names, for over three decades. Hence, the process of developing, testing and deploying these models is well-established in the industry.
BFSI organizations are already using digital twins in credit risk scoring, wealth management, customer life cycle management, fraud detection, and transaction monitoring. Most of these digital twins are focused on monitoring and predicting customer and portfolio key performance indicators such as creditworthiness, assets under management, customer loyalty, and customer lifetime value. However, few firms have considered how LLMs could streamline the analysis of unstructured data.
This is something that businesses across all domains were waiting for a way to finally leverage the huge amount of text data that was mostly unused. In the BFSI domain, LLMs are being tested for deployment as the next generation of analyzing and using text data. Early applications are, of course, chatbots and Q&A bots. Data augmentation, automation of reporting, automatic translation, and content generation for sales and marketing are expected to be the second generation of LLM applications.
Enterprises will need to develop new workflows to ensure the LLMs don’t hallucinate while structuring data into a format suitable for digital twins. Most digital twins are developed using time series data, which requires special storage and handling of the data in a serial manner. The bulk of enterprise data is stored in data warehouses based on relational databases. However, converting this data into the time series data required for digital twins usually involves additional data preparation steps requiring specialized software, data engineering skills, and storage formats.
Refaat also sees tremendous promise in using LLMs to improve the user experience for interacting with digital twins:
The presentation layer of digital twins will undergo a transformation over the next few years thanks to LLMs allowing a conversation-style interaction with all software applications. The recent popularity of using LLMs in question answering and generating content will drive innovation in digital twins to include natural language interactions in addition to existing charts, reports and dashboards.
The generative AI hype is driving many enterprises to explore ways of bolting on LLMs to their existing applications and services. Equal consideration must be given to how to structure the data upfront to reduce hallucinations and improve accuracy.
Early work with LLMs focused on how to train models to interpret text without regard to its provenance or how accurately it reflects the real world. Developing processes to structure this data into digital twins and mechanisms for tracking its provenance and meaning won’t be easy, but it is essential.
In the short run, this combination will help improve existing business processes. In the long run, it could make it easier to collaborate on more complex problems required to profitably meet sustainability goals.