Researchers and organizations often need to evaluate how complex systems might respond to new conditions before those conditions occur. For example, clinicians may want to evaluate how a new treatment might affect patient outcomes years in the future, even before running a clinical trial. Similarly, analysts and researchers may seek to understand how a financial market would react to a sudden policy shift, or how a power grid might behave during an unexpected heatwave.
Recent advances in generative AI (GenAI) make it possible to construct digital twins—high-fidelity virtual representations of real-world systems—that can simulate how those systems may evolve under different inputs or conditions.
At Analysis Group, our data science team is building digital twins powered by generative models that learn patterns from data in complex systems and generate simulated observations consistent with those patterns. These digital twins use time series and probabilistic models to simulate plausible trajectories under alternative inputs or conditions. These tools can help organizations explore potential outcomes across industries, from simulating patient outcomes to modeling financial resilience to forecasting energy dynamics.
Why This Matters
Organizations today operate in a world defined by uncertainty: Clinical trials can take years to produce results, market conditions can change rapidly, and critical infrastructure systems often cannot be safely stress-tested in the real world.
Historically, modelers tried to build these complex systems using equations, simulations, or statistical approximations. These approaches have been valuable, but often required restrictive simplifying assumptions.
That shift – from static models to dynamic GenAI-powered digital twins – makes it increasingly feasible to study complex systems under a wide range of scenarios. This framework allows organizations to evaluate many potential scenarios in a virtual environment before deciding which ones to test in practice. By digitally simulating outcomes, organizations can narrow down the most promising scenarios and prioritize those for real-world experimentation.
In essence: Analysis Group’s GenAI-powered digital twins don’t just describe systems, they simulate how those systems evolve.
GenAI Meets Digital Twins
We start with two key concepts:
- GenAI refers to models that can create synthetic data – such as text, images, or code – that reflect patterns observed in real-world data. They learn patterns and relationships from existing data, then use that understanding to generate new examples consistent with those patterns.
- A digital twin is a virtual counterpart to a real-world system, whether that’s a patient, a financial market, or a power network.
When you combine these two ideas, the result is a powerful analytical tool: A GenAI-powered digital twin that can simulate plausible future trajectories rather than simply reproducing past observations.
By learning from both fixed attributes (a patient’s blood type or a building’s design) and evolving ones (fluctuating weight, blood pressure, or energy demand), our GenAI models can simulate how systems behave over time and how they might respond under new conditions.
Historically, digital twins were built using mechanistic models based on domain expertise. While these models could capture some aspects of the system, they often struggled to represent the full complexity and variability of real-world dynamics. Now, with GenAI, digital twins can learn complex relationships directly from data without requiring all of those relationships to be specified in advance.
Inside the Model: Learning How Systems Evolve
Analysis Group develops GenAI digital twin models that learn relationships among relevant variables – from constant traits to those that change over time.
These models incorporate two main types of data:
- Time-invariant data that doesn’t change (think a patient’s blood type or a city’s latitude).
- Time-varying data that evolves from one moment to the next (think a person’s weight or a person’s daily energy usage). This includes variables that could be potential outcomes of interest, like cardiac risk.
Scroll through the visualization below to understand how these models learn the interactions between these elements.
During the first period, we provide the model with time-invariant data and baseline time-varying data.
We then feed these inputs into a GenAI algorithm, which
predicts the time-varying data for the next period.
That prediction then becomes part of the input for the next period, and so on.
Over many epochs (training rounds in which the model refines itself), the GenAI model can learn to generate simulated data that are statistically indistinguishable from real observations.
Importantly, the model doesn’t memorize data; it learns patterns and relationships present in the data. This means it can simulate scenarios that haven’t been observed yet but are consistent with the relationships learned from the data and domain expertise.
Training and Validation: Watching Twins Learn
During training, we can watch these GenAI-powered digital twins progressively learn the structure of the data.
Let’s look at an imagined example where we model patient health outcomes. Early on, the simulated data look random. Click through the slides below to see what we observe in the simulated data over time.
By the end, the simulated patients (or financial portfolios or power grids) can behave so realistically that even a sophisticated external machine learning classifier may struggle to tell them apart from real data.
Managing Model Limitations
While the GenAI-powered digital twins framework effectively learns and replicates complex relationships present in the observed data, it is inherently constrained by the scope and quality of that data. As a result, it may not capture patterns, variables, or dynamics that are absent from the original dataset. To address this, we design targeted evaluations to assess each model’s generalizability and ensure robust performance across diverse scenarios.
Exploring What-If Worlds
Once trained, the model gains a new capability: The GenAI powered digital twin becomes a flexible experimental environment.
By learning the relationships between all relevant variables in the system, the model’s use is not limited to a single research question or scenario. Without retraining the model to adapt to new questions, this framework can simulate a wide range of what-if scenarios, making it a versatile tool for forecasting and analysis.
Users can tweak inputs or focus on specific subgroups and instantly see projected outcomes. It’s like running 1,000 parallel worlds to see which ones lead to better futures.
Click through the examples below to explore how the GenAI powered digital twins framework can be applied to different practice areas.
Case in Point: Digital Patients at ISPOR 2025
At the ISPOR 2025 conference, Analysis Group showcased how GenAI-powered digital twins can revolutionize study of long-term disease progression and intervention effects. The challenge: Estimate how a 10% weight reduction might affect cardiovascular outcomes among obese patients, without waiting for years of trial data.
The solution: Our model simulated patient trajectories both with and without the weight loss intervention, demonstrating that GenAI powered digital twins can go beyond forecasting to bridge evidence gaps and generate high-confidence, timely insights into long-term outcomes.
Associated Contributors
For more information about the tools or technology in this post, please reach out to: Jimmy Royer (Principal, Montreal); Eric Wu (Managing Principal, Boston); Jinlin Song (Vice President, Los Angeles); Max Leroux (Director of Data Science, Montreal).
This GenAI methodology was also developed by: Intekhab Hossain (Senior Data Scientist, Boston); Yash Lalwani (Senior Data Scientist, Montreal); Esteban León (Senior Data Scientist, Boston); Lin Zou (Data Scientist, Boston); Varun Parkash (Senior Analyst, Los Angeles); Arina Rakhaeva (Senior Analyst, Paris).