Capturing Tablet Disintegration for Improved OSD Development

By Katie Anderson, Chief Editor, Pharmaceutical Online

Closed-loop laboratory automation promises to make oral solid dosage form development faster and more efficient, but the behavior of powders, tablets and dissolution processes remains difficult to predict. In a rapid fire presentation at this year’s PharmSci 360 event (Oct. 25-27, 2026), Pfizer senior scientist Hyunhong James Min will explain how his team is utilizing computer vision to measure disintegration kinetics for automated formulation and process development. I caught up with Min to get a sneak peak with his research in limitations of current dissolution models, the value of separating disintegration from API dissolution, and the steps needed to move the framework into an automated laboratory workflow.

What's the main bottleneck in closed-loop laboratory automation for oral solid dosage form development?

Hyunhong James Min (HJM): Two main bottlenecks are powder flowability and dissolution. Dissolution is particularly limiting because testing takes a long time, slowing iterative development and automation.

Why is dissolution testing particularly challenging to automate?

HJM: Going back to dissolution testing, one of the main issues is its relatively high variability.

Dissolution is affected by several factors, and formulation differences may be small, creating the compounding problems of high variability and small changes in mean response.

Statistical analysis may therefore find no significant difference even when meaningful effects exist. For example, smaller particles generally dissolve faster than larger particles, all else being equal.

Another challenge is that dissolution is a multifactorial problem. Trying to capture it using purely statistical methods, such as Bayesian optimization, can require a large number of experiments.

We have not yet explored this approach extensively, and I'm not even sure that having a large number of samples would necessarily solve the problem, given the variability in the results.

We therefore need a mechanistic tool that helps us understand how dissolution behaves.

Our mechanistic framework captures the aspects of dissolution we understand and uses a fitted factor for effects we do not fully understand.

What motivated you to separate tablet disintegration from API dissolution rather than modeling dissolution as a single overall process.  

HJM: We first characterized API powder dissolution and used it to predict tablet dissolution while accounting for disintegration. This works when the API behaves similarly in powder and tablet form.

That is not always true, particularly for poorly soluble drugs. Interactions within a tablet may affect API dispersibility or keep particles aggregated, so tablet disintegration must be distinguished from API dissolution.

We use the Noyes–Brunner equation to describe dissolution, but the equation alone does not fully explain the observed dissolution behavior because other factors also play a role.

After accounting separately for disintegration, we determine a fitted factor that represents API behavior within the tablet. Decoupling the two processes improves the model’s prediction of tablet dissolution.

What gaps in current dissolution models does this approach address?

HJM: The underlying framework has existed for nearly a century, but previous models often assumed a particular disintegration behavior instead of measuring its kinetics.

Our method uses computer vision to measure disintegration over time. We can then capture those kinetics and couple them to the dissolution model.

Rather than assuming that a tablet disintegrates in a particular way, we measure how it actually disintegrates and use that information in the model. We then determine a fitted factor that best describes the dissolution behavior of the API within the tablet.

We integrate measured disintegration kinetics with a dissolution model through population balance modelling, explicitly accounting for disintegration rather than treating the process as a single phenomenon.

How is the computer vision capturing disintegration kinetics from a time-stamped video? What features are you measuring?

HJM: Computer vision has improved considerably over the past decade. We use a pretrained model to track how the tablet's projected area changes over time.

As the tablet swells, wets and erodes, its projected area changes. The information about the tablet's disintegration and erosion is then fed into the population balance model.

One of the assumptions in our model is that, if 1% of the tablet has eroded based on the change in projected area, the corresponding material enters the dissolution pathway as primary API particles.

Those particles then dissolve into the dissolved state.

The fitted factor captures formulation and processing effects after initial disintegration. Because a tablet may pass through granules or other intermediate structures rather than directly becoming primary particles, the factor represents their combined effects.

How sensitive is the computer vision measurements to lighting, angle, etc.?

HJM: We use a zero-shot model developed by Meta called SAM [Segment Anything Model]. It has been trained on a broad range of images and works well under different lighting conditions. It can also identify objects when parts of them are obscured.

The precision of the measurements depends on the camera we use. Our camera is a microscope camera, with a resolution of approximately 24 micrometres per pixel. This gives us good spatial resolution.

We also monitor a robustness measure provided by the segmentation model, SAM 2. If that measure drops below a certain threshold, we manually check whether the segmentation is correct.

If the model incorrectly includes a tablet fragment in the main object, we adjust the prompts or segmentation in the code.

What are the differences between the fitting strategies, and why did you choose to evaluate each of those?

HJM: The first was to fit a constant factor, , to API powder dissolution data and then use that factor to predict tablet dissolution. As we discussed, this did not work well because API powder dissolution and tablet dissolution can behave quite differently.

The second strategy was to determine a constant after decoupling the disintegration kinetics from the dissolution process. This approach worked well.

The third strategy was to allow to vary with time. We explored five different terms or forms, including a constant value, an increasing factor and a linearly decreasing factor, with additional weighted terms.

The more flexible approach fitted the calibration data well, but performed worse when particle size changed and predictions were tested without refitting, suggesting overfitting.

What information becomes available When the disintegration is measured independently and decoupled from dissolution.

HJM: Many models assume that disintegration takes a fixed time and dissolution then begins, or that disintegration follows an exponential pattern. But disintegration is continuous, and our computer vision measurements show that its behavior can differ substantially.

During the initial period, typically the first 30 seconds or so, we may observe slight swelling or wetting. After that, the behavior depends on the tablet and its drug loading.

Some tablets undergo a sudden, almost explosive disintegration. Others undergo continuous, rapid erosion.

The disintegration mechanism therefore depends on how the tablet has been manufactured and formulated. It is not appropriate to assume that every tablet follows the same disintegration kinetics.

This information is increasingly important for polymer-containing formulations, where disintegration can be more challenging than with conventional excipients such as lactose and microcrystalline cellulose (MCC).

Some tablets are also highly sensitive to compression, particularly those with high drug loading.

In an example planned for a future paper, increasing tablet solid fraction from 0.8 to 0.9 raised USP disintegration time from about one minute to 11 minutes and dramatically slowed dissolution.

How well did the model predict dissolution across API lots with different, vertical size distributions.

HJM: The model predicted dissolution profiles well. For smaller particle sizes, predictions were generally within about 10 % of measured values. For larger particles, a measured dissolution of 50% corresponded to a predicted range of roughly 45%–55%.

How would Bayesian optimization use this factor to select the next formulation? Condition or process conditions.

HJM: We did not directly optimize the formulation or process; we demonstrated a framework using an uncoated core and tablets with different coating levels, including 1% and 5%, because coating was known to affect dissolution.

We calculated the disintegration kinetics and determined the corresponding value for the uncoated core and the 5% coated tablet. Because the API particle size, molecule and other relevant properties were the same, we could examine the effect of coating on the fitted factor.

We observed a reduction in of approximately 22% at the higher coating level. This captured the impact of coating on the dissolution behaviour.

A Gaussian Process can predict the factor at intermediate coating levels. With only two experimental points, we might assume a simple relationship; additional observations support a more informative model.

We then predicted the factor at an intermediate coating level, measured the disintegration kinetics and compared the resulting predictions with the measured data. The predictions agreed well with the experimental results, with the points falling within the reported standard deviation.

The approach will not work for every formulation, but it provides a framework combining computer vision, measured disintegration kinetics, population balance modelling and a dissolution model to determine the fitted factor.

We can then use a Gaussian Process to model how formulation and process variables affect that factor, making it possible to incorporate the relationship into a Bayesian optimization workflow.

Ultimately, if we understand how variables such as particle size, formulation and processing affect the fitted factor, we may be able to reduce the number of experiments required during formulation development.

How much experimental time or dissolution testing capacity could this approach save?

HJM: Additional experiments are needed to quantify savings compared with a traditional approach. The framework’s value is separating formulation and processing effects from underlying dissolution behavior.

If dissolution is strongly limited by disintegration, the model may be able to predict that behavior. When we develop batches of variant tablets, disintegration is one of the first things we examine.

We can use the model to predict which tablets are likely to show particular dissolution behaviour and identify where additional experiments are needed.

For example, we could test normal and over-lubricated tablets, predict behavior at an intermediate lubrication level, and decide whether that condition warrants testing rather than evaluating every combination.

I cannot give a definitive figure for the time or testing capacity saved without further validation. However, I would estimate that the potential reduction in experimental effort could be around 50%, although that would need to be confirmed experimentally.

What level of accuracy or uncertainty calibration is required before deploying the model in an automated laboratory?

HJM: That's a very good question, and it is not something I can answer with a single number.

As we gather more data, the model should improve its understanding of the relationships between formulation, process variables and dissolution.

Ideally, the model's predictions would fall within the experimental variability of the dissolution measurements. However, I do not believe we are at that point yet.

As we collect more data, we will develop a stronger mechanistic understanding of which molecular and formulation characteristics affect dissolution, and which process factors are important.

We will gradually build up our prior knowledge. This should reduce the number of experiments needed to identify an optimal formulation or process condition.

Before automated deployment, we need more data and validation. More observations and stronger prior knowledge should reduce the experiments required to reach an optimal state.

What remains to be accomplished in this research?

HJM: This is only the beginning. The immediate next step is integrating the framework into a laboratory automation workflow and demonstrating it in practice.

As experiments accumulate, we want to strengthen our understanding of relationships among formulation, processing and dissolution, reducing the tablets and experiments needed to identify optimal conditions.

The computer vision method also needs further work. It currently measures two-dimensional projected area and assumes uniform surface erosion, which is not always accurate.

Erosion may vary with tablet orientation and exposed surfaces, so projected area may not represent the actual remaining volume.

A three-dimensional approach using multiple camera angles could capture tablet geometry and disintegration more accurately, helping refine the model.

The Framework For Future Formulation

The work remains at an early stage, but it points toward a more informed and selective approach to formulation development. By linking direct observations of tablet disintegration with mechanistic modelling and adaptive experimental design, the framework could help researchers identify the most valuable experiments, reduce unnecessary testing, and better understand how formulation and processing choices affect dissolution. Further validation, integration into automated workflows, and improved three-dimensional measurements will determine how broadly the approach can be applied.