Colloidal quantum dots (QDs), also known as colloidal semiconductor nanocrystals (NCs), are zero-dimensional nanoscale semiconductor particles (typically 2–10 nm in size) whose electronic and optical properties are size-dependent due to their quantum confinement [1]. Their size-dependent optical properties have been the focus of significant research over the past two decades [2]. Each QD consists of between a few hundred to a few thousand atoms [3], and surrounded by an outer layer of functional groups such as amino, carbonyl, aldehyde, hydroxyl, and carboxylic acid groups [4]. In a system when the particle radius becomes smaller than the exciton Bohr radius, quantum confinement renders the bandgap size-dependent, allowing for precise tuning of absorption and emission in the visible spectrum [5]. Consequently, changing the size of nanocrystals by a few hundred atoms can shift their emission from the deepest red to the vibrant blue, which is a concept of incomparable power. For any material, the ratio of surface area to volume increases as its size decreases. As the dimensions reduce, surface effects become more significant and eventually dominate over the bulk properties [6]. Their solution processability, photoluminescence tunability, and narrow emission make them key to a wide range of applications, including drug delivery, energy storage, photovoltaics, displays (QLEDs) [7], [8], [9], [10], [11], solar cells [12], photodiodes [13], photoconductors [14], bioimaging and sensing [15], field effect transistors [16] and future quantum information devices [17]. The size-tunable fluorescence is the hallmark property of QDs, enabling their vibrant and precisely controllable color palette for diverse applications [6], [18]. The multi-dimensional parameter spaces (precursors, temperature, ligands, reaction time, doping, surface chemistry) and complex growth physics make the problem an expensive, time-consuming experimental search problem [19]. The high surface area to volume ratio means a significant fraction of atoms at the interface [20], presenting coordination unsaturation, and these atomic vacancies dangling bond act as trap states for charge carriers, creating mid-gap electronic states, acting as a non-recombination center, quenching photoluminescence, and compromising device performance [21], [22]. Machine learning models analyze high-dimensional datasets linking conditions, chemical compositions, and structural features with desired optical, catalytic, biomedical, electronic energy, and sensing applications [23]. The recent awarding of the 2023 Nobel Prize in Chemistry for the discovery and synthesis of QDs has cemented their status as a foundation of modern nanoscience [24].
Despite this immense and now notable promise, a significant challenge has persistently hindered their progress from the laboratory to global application: the synthesis itself. The synthesis of a perfect, monodisperse quantum dot is not a straightforward path but goes through a vast chemical space. Minute shifts in precursor concentrations, ligand identities and ratios, reaction temperature, injection speed, reaction viscosity, and aging times can drastically alter growth kinetics and the resulting shape and size distribution, often leading to failed experiments [25], [26], [27]. Navigating this space demands precise control over multiple factors to avoid unwanted phase transitions or polydispersity. Traditionally, the design and synthesis of QDs relied on trial, error, and serendipity, which can consume months or even years of dedicated research to optimize a single material [28]. This problem has defined the central bottleneck in the field. The research in modern era, driven by data-intensive science. The integration of artificial intelligence (AI) and machine learning (ML), deep learning (DL), and generative models offers a powerful approach to explore these spaces, optimize processing and synthesis, and predict properties from data of multiple precursor combinations as input variables [29], [30]. In contrast to these traditional empirical methods, AI-assisted approaches offer an inexpensive and accelerated approach for the more accurate design and discovery of new materials, particularly for clean energy applications [31]. AI not only enabled the accelerated prediction of various structural property relationships, automated discovery, optimization, and forecasting, but also led to a deepened understanding of structures and precise properties of a QD, including its emission peak to within a nanometer and its quantum yield to a percentage point, before a single expensive precursor is ever used [28].
A comprehensive bibliometric analysis reveals that the term “Quantum Dot” appears in 1,63,472 publications. In contrast, only 1,522 publications are found when the search is narrowed to include the intersection of “Quantum Dot” with terms related to artificial intelligence, specifically “Artificial Intelligence,” “Machine Learning,” “AI,” or “ML” keywords [2].
This analysis shows that research on QDs was growing in the early 1990s and accelerated after 2005. The high number of publications in recent years (now exceeding 10,000 annually) demonstrates continued strong scientific interest in QDs studies (Fig. 1A). In the field of QDs and AI/ML, meaningful growth has been seen since 2006, and in the last seven years, its application has been exceptionally increased, indicating a powerful convergence of the field. The total number of publications (1,522) is small relative to the overall quantum dot literature, confirming that this is a new and emerging research boundary (Fig. 1B). The overall research volume in Fig. 1A is over 100 times larger than in Fig. 1B, but the relative growth rate in recent years appears to be much higher in Fig. 1B. This suggests the application of AI/ML to QDs is one of the fastest-growing areas within multiple fields and applications due to the ability to accurately predict the design and synthesis processes.
Carbon quantum dots (CQDs), first discovered in 2004 during the purification of single-walled carbon nanotubes [32], are well known for their photoluminescent, electroluminescent properties, along with fast response time and excellent sensing performance [33]. These attributes make QDs highly significant for applications in industrial research. However, the main challenge in their preparation is to produce a nanomaterial with a high quantum yield (QY) that ensures its high sensitivity and reliability. Historically, the widespread use of the trial-and-error approaches has been time-consuming and resource-intensive, generating considerable wastage [34]. The traditional method encounters substantial obstacles due to the extensive range of reaction parameters and variability in precursors [35], [36]. To overcome these limitations, researchers are now implementing various computational models of AI and ML for synthesis with desired properties [37]. Over the past decade, ML has emerged as a strong tool due to its exceptional predictive capabilities. The integration of AI and ML with experimental platforms and various computational tools is transforming the field of chemistry by enhancing predictive accuracy and supporting human decision-making. AI-assisted material design and discovery methods can facilitate material development rapidly as well as inexpensively [31]. These methods mainly aim to optimize nanomaterial properties through efficient exploration of their synthetic parameter space and guide the smart fabrication with customized properties of QDs synthesis and application. This whole ML process can be concluded in steps of optimization of properties, creation of programmable nanoparticles, engineering of nanostructure surface, elucidation of complex signals, and then construction of accurate prediction models [38].
Machine learning excels at identifying patterns from large datasets, often referred to as ‘Big Data’, to solve problems and deliver rapid, precise predictions through simulations [39]. As a result, AI-based research efforts in the synthesis of various nanomaterials, including QDs, is growing exponentially to identify and produce materials with targeted properties. For example, Krishnadasan et al. created an autonomous ‘black-box’ system by using a stable noisy optimization by branch and fit (i.e., SNOBFIT) to find and manage the optimal injection rate for the synthesis of CdSe QDs within a microfluidic reactor [40]. Similarly, Epps and co-workers introduced an ‘artificial chemist’ for machine learning based experiment selection with an automated flow reactor. This setup was used to fine-tune and optimize the fluorescence properties along with the photoluminescence quantum yield (PLQY) of CsPb(X)3 QDs, where X represents either bromine or iodine [35].
But prediction, however accurate, is only the beginning. The true revolution lies in inverse design: the ability to invert the scientific process itself [41]. Instead of asking “What properties will this material exhibit?”, inverse design enables researchers to ask “What material should I make to get the properties that I need?”. One can begin with a thought for property, a specific color for a light emitting diode (LED) screen, a target PLQY for a solar concentrator, enhanced stability for a biological probe, and task the AI to design not only the novel material but also the detailed synthetic recipe to create it in the laboratory. For example, in 2023, Chen and team utilized ML model to improve the operational stability of blue quantum dot light emitting diodes (QLEDs), a critical barrier to commercialization [42]. From more than 200 samples, they fabricated over 800 blue QLEDs by the use of convolutional neural network and achieved a Pearson correlation coefficient of 0.70 in the test of lifetime prediction. The fusion of AI prediction and autonomous synthesis for application-oriented optimization marks a transformative period for QD technology, with the potential to drastically reduce discovery timelines and unlock entirely new classes of functional nanomaterials.
In this review, we focus exclusively on the role of AI and ML techniques in the design and application process of different QDs. We first discuss the fundamentals of AI with QDs and their structure-property relationships, followed by methodologies enabling accelerated materials discovery. The subsequent section will deal with detailed AI applications in different fields like optoelectronics, sustainable energy, biomedical engineering, and sensing. Finally, we assess current limitations and outline future opportunities in the field. We anticipate that this review will serve as a valuable reference for the broader research community in the field and offer insights into the design and development advancements in AI-driven synthesis techniques.
In 1965, Nelder and Mead introduced the simplex algorithm, the first published article providing a foundational optimization method [43]. This landmark optimization technique, now known as the Nelder-Mead method, remains widely used today, especially in chemistry, engineering, and machine learning when derivative information is unavailable [44]. QDs were first conceptualized in the early 1980s when physicists recognized that semiconductor nanocrystals could confine electrons in their dimensions, revealing that nanoparticles could exhibit electronic properties resembling discrete atomic states. Louis E. Brus at Bell Labs and Alexei Ekimov in the Vavilov State Optical Institute (Petersburg, Russia) reported the synthesis of semiconductor nanocrystals that exhibited size-dependent optical properties. In 1981, Alexei Ekimov was the first to document how the optical properties of light varied with particle size [45]. In the subsequent year, Brus reported synthesizing QDs in a solution as part of his investigation into semiconducting particles for their solar energy applications [46]. These early studies demonstrated that reducing the size of the semiconductor to the nanoscale induced quantum confinement effects, resulting in tunable emission spectra.
During the 1990s, there was a rapid advancement in colloidal synthesis methods, allowing researchers to produce QDs with desired controlled size and enhanced photoluminescence efficiency. A milestone was Moungi Bawendi’s 1993 development of highly controlled and precise synthesis protocol for QDs combining inorganic and organometallic chemistry [25]. This enabled the size-controlled synthesis of nanoparticles with their tunable optical properties. This period established the foundational understanding of QDs as versatile materials for optoelectronics, photovoltaics, imaging, electronics, and display technologies.
The decade of the 2000s was characterized by refinements in chemical synthesis, particularly hot-injection methods, which enabled the production of highly monodispersed QDs with narrow emission spectra. Yu et al. successfully synthesized CdS nanocrystals using non-coordinating solvents (specifically octadecene) and oleic acid as ligand [47]. During this time, cadmium-based QDs (e.g., CdSe, CdTe, etc.) became widely studied and commercially applied in the fabrication of multicolored light-emitting displays [48], lighting, solar cells (photovoltaics), and early biomedical imaging and (bio)sensor tools [49], [50], [51], [52], [53], [54], [55]. Researchers started developing intelligent routes for the controlled synthesis of nanoparticles in microfluidic reactors using different algorithms. However, toxicity concerns regarding lead and cadmium prompted interest in lead-free alternatives like carbon-based QDs [56]. Meanwhile, researchers began integrating computational modeling to predict QD’s behavior, although these models were still limited by computational power and the complexity of nanoscale systems.
In the decade of 2010, Pan et al. synthesized graphene quantum dots (GQDs) by using graphene for the first time, which have unique properties as high biocompatibility, low toxicity, excellent chemical stability, and tunable photoluminescence. GQDs exhibited strong quantum confinement due to their small size and edge structure [57]. This decade can be marked as a crucial transition with the rise of AI and ML as tools for materials science and accelerated discovery through ML and robotics [58], [59]. High-throughput experiments began generating large datasets on QD synthesis parameters, and optical datasets on QD synthesis parameters. ML models, like the support vector machine (SVM) [60] and random forests (RF) [36], were applied to identify patterns within these datasets, enabling more systematic optimization of the properties of QDs. In 2017, a significant development was the use of natural language processing (NLP) and ML for automatic compilation of materials synthesis parameters from scientific literature, allowing the prediction of critical parameters for materials like titania nanotubes via the hydrothermal method [61]. Deep learning further expanded capabilities by predicting nonlinear relationships between synthesis conditions and enhanced resulting material properties with broadening their applications. Similar developments in perovskite QDs introduced other new challenges, such as instability under moisture and light exposure, which were addressed by the help of AI tools suggesting alternative compositions and passivation approaches [62].
In the 2020 s decade, Han et al. developed an extreme gradient boosting (XGBoost) machine learning model to reveal the relationship between various synthesis parameters and experimental outcomes, which led to the successful creation of green-emissive carbon dots (CDs) with enhanced fluorescent QY of up to 39.3 % [60]. Due to high sensitivity and low cost, different types of QDs have started to be used for sensing purposes of pesticides [63], [64], [65], [66], [67], bio enzymes [68], [69], heavy metal ions [68], [70], [71], [72], glucose [73], [74], and other important analytes [75], [76], [77]. Building on this, AI-driven approaches moved from predictive modeling to generative design. Reinforcement learning, which learns repetitively through trial-and-error datasets, began to propose novel synthetic pathways by measuring its progress towards an individual goal. Neural networks were used to optimize QD emission wavelength for display technologies and bioimaging applications [78], [79]. Generative adversarial networks (GANs) [80] utilize two competing networks: a generator that creates new data and a discriminator that evaluates its authenticity against a real dataset. Autoencoders are known for their use in designing molecules through training to contain an individual property; variational autoencoders (VAEs) enabled the discovery of new QD compositions, accelerating the search for non-/low toxicity, stable materials with better properties suitable for commercialization. AI is now incorporated with robotic systems, performing high-throughput synthesis and characterization of QDs under the guidance of machine learning models. The time needed to find the best material configurations is significantly decreased by this type of closed-loop method. Numerous industries have changed as a result of the merging of AI and QD research, which is increasing in strength every day.
In this review, we present a comprehensive and critical perspective to the rapidly growing limits, aiming to separate the noticeable successes from the projected hype. We begin by establishing foundational concepts, outlining the fundamental principles of both QDs and the AI toolkits deployed to study them. We then examine the AI-driven workflow for prediction, optimization, and inverse design to implement in practice in order to create QDs with enhanced features and applications. Next, we will assess the impact of these AI-engineered materials having across the most significant technological applications, asking where this new approach has already made a measurable difference. Finally, we critically address the significant challenges that remain from the hard truth of data scarcity to the ‘black box’ nature of AI and offer our viewpoint on the future for this exciting human-AI collaboration in materials science.