Active learning in pharmaceutical 3D printing: a multi-dataset comparison

Abstract

Machine learning (ML) is expected to accelerate the developments of three-dimensional (3D) printed medicines. Despite ML’s potential, the need for large datasets can hinder progression, as 3D printing remains an emerging pharmaceutical manufacturing technology. This study explores an ML strategy called active learning (AL), which harnesses the benefits of ML whilst applicable with small datasets. AL was tested to predict the printability of three 3D printing datasets: 1437 fused deposition modelling (FDM), 650 vat polymerisation and 297 selective laser sintering (SLS) formulations. The analysis revealed that accuracies of 60% can be achieved when starting with 33 formulations, and subsequent increases in training data size enhances predictive performance. Furthermore, AL was found to achieve 100% predictive accuracy, which is the highest recorded to date for pharmaceutical 3D printing. These initial findings highlight AL’s advantages over traditional ML modelling and showcase its potential to accelerate the development of 3D printing medicines. This research also demonstrates the potential of modelling with small datasets, thereby widening ML’s application in pharmaceutical research.

Introduction

Three-dimensional (3D) printing, or additive manufacturing, is a transformative technology in the field of medicine manufacturing [1,2,3,4,5,6,7]. The technology produces novel and innovative medicines that have the potential to address unmet clinical needs, particularly in precision medicine [8,9,10,11,12,13]. Recent examples include Spritam®, a 3D printed medicine for epilepsy, and the recent clearance by the food and drug administration (FDA) for several of Triastek’s investigational new drugs [14, 15]. Briefly, the technology provides an unprecedented level of spatial control, whereby both the interior and exterior structure of a medicine can be precisely designed. This added level of sophistication provides an opportunity to conform to the precision medicine paradigm. 3D printing is advancing our ability to precisely control the dose and drug release profile, such that medicines can be tailored to meet individual needs [16,17,18,19,20].

It is evident that 3D printing has great potential. However, since Spritam’s FDA approval in 2015, there has been little bench-to-bedside translation, which is expected of any nascent technology [21, 22]. Furthermore, as 3D printing can produce near-limitless designs, while indeed a great benefit, means substantial and laborious research is needed to explore all possible designs. Recently, numerous modelling approaches have accelerated developments. This includes physics-based computational models, design of experiments (DoE) and mechanistic models [23,24,25,26,27,28]. Mixture designs, a type of DoE, can be implemented to consider the formulation space [29,30,31,32], but cannot rapidly model it. Like most DoEs, the modelling will function only when all the requested experimental runs have been completed, making the process time-consuming. For example, a mixture design with 10 components (i.e., ingredients) can result in over 1000 experimental runs. Hence, most mixture designs reported in the literature are limited to less than 10 components [33,34,35,36,37].

More recently, machine learning (ML), a subset of artificial intelligence (AI), has been applied to accelerate developments. ML is an emerging modelling technology that has achieved remarkable feats in recent years and transformed the field. In pharmaceutics, ML has been applied across the entire workflow [38, 39], from upstream processes such as predicting drug-excipient compatibility and formulation processability [40,41,42], to predicting both in vitro and in vivo outcomes [43,44,45]. Such studies have demonstrated the capabilities of ML modelling in accelerating developments in pharmaceutics. In pharmaceutical 3D printing, research has revealed ML’s utility in tackling numerous processing issues. For quality control, ML has been found to accurately classify the structural quality of 3D printed products in real-time [46, 47]. Moreover, ML has been reported to not only identify the active pharmaceutical ingredient (API) in a formulation but also verify its dose [48]. Furthermore, several studies have demonstrated that ML can predict the 3D printability of formulations and has thus far been applied to fused deposition modelling (FDM), selective laser sintering (SLS), digital light processing and inkjet printing [49,50,51,52,53,54]. A cutting-edge application of ML in pharmaceutics is the use of generative AI to predict potential formulations [55,56,57,58,59], and by combining AI’s ability with 3D printing, a novel recent study has enabled the generation of de novo 3D printed formulations, and found ML could generate novel formulations that captured the qualities of human-generated formulations [60]. Collectively, these studies have demonstrated the advantages of leveraging ML for pharmaceutical 3D printing.

The benefits of ML over other computational models include the ability to comprehend complex and high dimensional data, interpret different data formats (e.g., numeric, text, images) and super-human comprehension speeds [61,62,63,64]. The high dimensionality allows ML to analyse a large number of components simultaneously without needing additional, exhaustive trials. Together, these aspects allow ML to model pharmaceutical processes that have proven challenging with traditional models. Previous work in the field of pharmaceutical 3D printing include predicting the printability and processing temperature, the drug loading and the drug release rate [48, 65, 66]. However, a prominent disadvantage of ML, is the need for large datasets, resulting in researchers and clinicians alike needing to amass numerous formulations to harness the benefits of ML. For example, a recent study achieved prediction accuracies above 90% in predicting printability but required mining the literature for over 900 formulations [66]. This limits the application of ML to the latter stages of development, preventing implementation at the early stages.

The need for large datasets in ML is not exclusive to pharmaceutical sciences. The cost and time associated with generating more data afflicts allied fields and beyond. This has attracted a concerted effort by researchers to develop ML models for small datasets. One particular field of interest is active ML or active learning (AL). It is termed ‘active’ because it actively learns as it receives more data, which helps to improve its predictive performance. AL models continuously interact with researchers throughout the experimental process and, remarkably, can dictate the experiments conducted for its learning. AL’s application in guiding researchers through the experimental space allows AL to not only make predictions but also alleviate the decision making of which experiments to test next. This makes AL useful in the early stages of development, as it can be trained on a small dataset yet tasked with predicting outcomes for much larger datasets. Simply put, the primary goal of AL is to map the experimental space with as little experiments as possible. In molecular medicine, several studies have found that AL can drastically reduce the number of experiments conducted [67,68,69,70,71,72,73,74,75]. In addition to its accuracy, the speed at which AL performs has prompted researchers to incorporate AL into automated laboratory robots, such that AL guides the robotic platform in which samples to test, rather than testing all samples [76].

To that end, we investigated the potential of AL for accelerating developments in pharmaceutical 3D printing. To examine its versatility, we applied AL retrospectively to three distinct datasets (fused deposition modelling (FDM), vat polymerisation and selective laser sintering (SLS)), where each one was comprised of different sample sizes, numbers of components and ratios of printable to non-printable formulations. It is worth noting that this was the first study in predicting the printability of vat polymerisation pharmaceutical formulations using ML. We examined the effects of starting sample size and the effects of three different machine learners. The principles of AL are provided in the supplementary (Section S1).

Download the full article as PDF here Active learning in pharmaceutical 3D printing

or continue reading here

Elbadawi, M., Kafoor, N.F.A., Li, H. et al. Active learning in pharmaceutical 3D printing: a multi-dataset comparison. Drug Deliv. and Transl. Res. (2026). https://doi.org/10.1007/s13346-026-02077-x