Colloidal quantum dots have earned their place at the center of modern nanotechnology, powering everything from television displays to solar cells and biomedical imaging agents. Yet for all the attention devoted to their dazzling size-tunable optical properties, the true gatekeepers of their performance lie at the nanocrystal surface, where a forest of organic ligands binds to the inorganic core. How exactly these molecules attach, contort, and transform on the facets of a quantum dot has long been too complex to observe directly in simulation. Now, a team at Zhejiang University has changed that, using machine learning to build a force field capable of simulating large cadmium selenide quantum dots in full atomic detail, and in doing so revealing a hidden taxonomy of ligand geometries that could shape how researchers design future quantum dot materials.
The study, led by Haibing Zhang, Bichuan Cao, Lei Huang, Xiaogang Peng, and Linjun Wang of the Department of Chemistry at Zhejiang University, was published in Nano Research. Its central achievement is a machine learning force field, or MLFF, built on the DeePMD framework, that makes molecular dynamics simulations of realistically sized quantum dots computationally feasible without sacrificing the accuracy of quantum mechanical calculations. The work addresses a stubborn bottleneck in computational nanoscience: quantum dots large enough to matter experimentally contain thousands of atoms, far beyond the reach of conventional first-principles molecular dynamics, while classical force fields lack the flexibility to describe the messy, fluxional chemistry of a ligand-capped nanocrystal surface.
Machine learning force fields have emerged over the past decade as a way to split the difference. The underlying idea, traceable to the landmark neural network approach of Behler and Parrinello in 2007, is to train a neural network on a library of quantum mechanical calculations so that the network learns to predict the energy of each atom from its local atomic environment. Once trained, the model evaluates energies and forces at a tiny fraction of the cost of a density functional theory calculation, yet reproduces the quantum mechanical potential energy surface with remarkable fidelity. The DeePMD framework extends this idea with deep learning architectures that respect the symmetries of physics, ensuring that the predicted energy does not depend on how the molecule is rotated or translated in space.
What makes the Zhejiang team’s approach distinctive is the way they constructed the training data. Building a reliable force field for quantum dots requires examples that capture every relevant bonding motif, coordination environment, and surface reconstruction the simulation might encounter. Rather than laboriously hand-picking configurations, the researchers devised an effective strategy to generate a diverse dataset of small quantum dots, where quantum mechanical reference calculations remain affordable. Small clusters of zinc-blende CdSe, capped with carboxylate ligands, were sampled across a broad range of structural distortions and binding scenarios. Using the DP-GEN concurrent learning platform, the dataset was iteratively expanded to cover regions of configuration space where the model’s predictions were least certain, a technique that concentrates computational effort precisely where the force field needs more training.
The resulting model, which the authors call the QD force field, or QDFF, expresses the energy of each atom based purely on its local atomic structure. This locality assumption is what allows the force field, trained on small quantum dots, to be transferred to much larger ones. Once validated, the team deployed QDFF to run molecular dynamics simulations of large zinc-blende CdSe quantum dots passivated with carboxylate ligands, the workhorse ligand class used in colloidal synthesis. Carboxylates, with their negatively charged head group anchoring to surface cadmium ions and their hydrocarbon tails extending into solution, are the molecules that keep quantum dots dispersed, stable, and processable. Their arrangement on the surface, however, has been notoriously difficult to pin down.
The simulations revealed a surprisingly rich picture. The researchers identified four major binding geometries for carboxylate ligands on the CdSe surface: bridging, in which the ligand’s two oxygen atoms attach to two different cadmium sites; tilted, in which the carboxylate group leans away from a perpendicular orientation; chelating, in which both oxygens bind to a single cadmium atom; and claw, a more intricate arrangement in which the ligand appears to grip the surface in a pincer-like fashion. None of these geometries is static. Over the course of the simulations, ligands were observed migrating between binding modes, and the frequency of each geometry depended strongly on the length of the alkyl chain attached to the carboxylate head. Longer chains, it turns out, do more than just push neighboring dots apart in solution; they reshape the local binding landscape on the surface itself.
To go beyond a static catalog of geometries, the team turned to Markov state models, a statistical framework borrowed largely from protein folding studies that dissects long molecular dynamics trajectories into discrete states and the transition pathways connecting them. Applied to the ligand dynamics on the quantum dot surface, the Markov analysis exposed the detailed channels through which carboxylate ligands transform from one geometry to another, effectively producing a kinetic map of the surface chemistry. This level of detail matters because ligand geometry is not merely structural trivia. The way a ligand binds influences the electronic structure of the nanocrystal, its optical spectra, and its exciton dynamics, including the notorious blinking behavior and hole trapping processes that plague quantum dot emitters. Prior experimental work from the same group had already shown that carboxylate coordination structures on CdSe nanocrystals depend on the crystal facet, and that facet-ligand pairing can drive reversible surface reconstruction.
The significance of the new work lies in the marriage of scale and fidelity. Earlier computational studies of quantum dot surfaces were restricted to small clusters or short timescales, or relied on classical potentials that could not describe the breaking and reforming of coordination bonds. Ab initio molecular dynamics, the gold standard for accuracy, becomes prohibitively expensive for systems containing several thousand atoms and trajectories long enough to capture rare events. QDFF sidesteps both limitations, offering near-quantum accuracy at a computational cost that makes extended simulations of large, fully ligand-passivated quantum dots routine. The authors note that the high performance of the force field makes the approach promising for systematic studies of large quantum dots decorated with the many different ligand chemistries that can be synthesized in the laboratory.
The implications reach across the quantum dot field. Device engineers have long known that surface passivation quality governs the performance of quantum dot light-emitting diodes and solar cells; charge leakage, trap formation, and non-radiative recombination all trace back to imperfect surface chemistry. Ligand-related surface traps have been shown to be mobile on nanocrystal surfaces, and the mobility observed in the new simulations provides a dynamic, atomistic picture of exactly how such traps might form, migrate, and heal. By identifying which ligand geometries dominate and how they interconvert, the study gives synthetic chemists a concrete target: if particular binding modes correlate with electronic traps or degraded luminescence, ligand design can aim to suppress those modes, whether by adjusting chain length, branching, or anchoring chemistry.
There is also a methodological lesson in the work. The strategy of training on small, fully quantum mechanically characterized systems and transferring to large ones could be generalized well beyond CdSe carboxylate systems. Core-shell nanocrystals, alloyed quantum dots, perovskite nanocrystals, and other ligand-capped semiconductor systems all face the same computational wall. A validated QDFF-style force field, retrained for a new material system, could open similar windows onto surface dynamics that have remained opaque. The combination of machine learning force fields with Markov state modeling, in particular, offers a general recipe for extracting kinetic information, not just structural snapshots, from large-scale simulations of nanomaterials.
The research, supported by the National Natural Science Foundation of China, reflects a broader convergence of artificial intelligence and molecular simulation that is reshaping computational chemistry. As neural network potentials grow more accurate and the platforms for generating their training data more automated, the class of systems amenable to realistic atomistic simulation keeps expanding. For quantum dots, whose commercial success rests on exquisite control of surfaces, the arrival of a tool that finally renders those surfaces visible in motion may prove to be a turning point. The tiny ligands draped over a nanocrystal, once invisible to theory, are now coming into sharp focus, and with them the atomic-scale levers that determine whether a quantum dot shines brightly or flickers uselessly in a device.
Subject of Research: Machine learning force field simulation of carboxylate ligand geometries and dynamics on the surface of zinc-blende CdSe quantum dots
Subject of Research: Technology and Engineering
Article Title: Machine learning force field study of carboxylate ligands on the surface of zinc-blende CdSe quantum dots
Article References: Zhang, H., Cao, B., Huang, L., Peng, X., & Wang, L. (2024). Machine learning force field study of carboxylate ligands on the surface of zinc-blende CdSe quantum dots. Nano Research, 17(12), 10685-10693. https://doi.org/10.1007/s12274-024-6983-9
Image Credits: AI Generated
DOI: 10.1007/s12274-024-6983-9
Keywords: quantum dots, ligand geometry, machine learning force field, molecular dynamics, CdSe nanocrystals, carboxylate ligands, DeePMD, Markov state model, surface chemistry, zinc-blende
Cite Scienmag News
APA MLA Chicago
Blake Davidson. (September 5, 2026). Machine learning force fields reveal carboxylate ligand binding on CdSe quantum dots. Scienmag. https://scienmag.com/machine-learning-force-fields-reveal-carboxylate-ligand-binding-on-cdse-quantum-dots/
Blake Davidson. “Machine learning force fields reveal carboxylate ligand binding on CdSe quantum dots.” Scienmag, 5 September 2026, https://scienmag.com/machine-learning-force-fields-reveal-carboxylate-ligand-binding-on-cdse-quantum-dots/. Accessed 5 September 2026.
Blake Davidson. “Machine learning force fields reveal carboxylate ligand binding on CdSe quantum dots.” Scienmag. September 5, 2026. https://scienmag.com/machine-learning-force-fields-reveal-carboxylate-ligand-binding-on-cdse-quantum-dots/
Copy citation Download RIS
Tags: advanced simulation techniques for nanomaterialsadvanced simulation techniques in nanomaterialsatomistic modeling of colloidal quantum dotsatomistic modeling of quantum dot surfacescadmium selenide quantum dotscomputational design of quantum materialsDeePMD framework for molecular simulationsDeePMD-based machine learning modelsdesigning functional quantum dot materialshidden ligand geometries on quantum dotsligand binding mechanisms in nanomaterialsligand binding on CdSe quantum dotsligand geometry and quantum dot performancemachine learning force fieldsmolecular dynamics simulation of nanocrystalsmolecular dynamics simulation of nanomaterialsnanotechnology applications in displays and solar cellsnanotechnology in biomedical imagingorganic ligand conformations on nanocrystalsorganic ligand interactions with quantum dotsquantum dot surface chemistrysize-tunable optical properties of quantum dots
