Nanoscience News (<front>)

Certain about uncertainty? Latent representations of VAEs optimized for visual tasks.

Deep Learning methods are increasingly becoming instrumental as modeling tools in Computational Neuroscience, employing optimality principles to build bridges between neural responses and perception or behavior. Deep Generative Models (DGMs) can learn flexible latent variable representations of images while avoiding intractable computations, common in Bayesian inference. However, investigating the properties of inference in Variational Autoencoders (VAEs), a major class of DGMs, reveals severe problems in their uncertainty representations. Here we draw inspiration from classical computer vision to introduce an inductive bias into the VAE by incorporating a global explaining-away latent variable, which remedies defective inference in VAEs. Unlike standard VAEs, the Explaining-Away VAE (EA-VAE) provides uncertainty estimates that align with normative requirements across a wide spectrum of perceptual tasks, including image corruption, interpolation, and out-of-distribution detection. We find that restored inference capabilities are delivered by developing a motif in the inference network (the encoder) which is widespread in biological neural networks: divisive normalization. Our results establish EA-VAEs as reliable tools to perform inference under deep generative models with appropriate estimates of uncertainty.

• Speaker: Josefina Catoni, Universidad Nacional del Litoral
• Wednesday 12 February 2025, 15:00-16:30
• Venue: CBL Seminar Room, Engineering Department, 4th floor Baker building.
• Series: Computational Neuroscience; organiser: Daniel Kornai.

Add to your calendar or Include in your list

Inverse normative modeling of continuous perception and action

Normative models of behavior strive to explain why behavior unfolds the way it does and have been highly successful in explaining many phenomena in neuroscience, cognitive science, and related fields. The power of these approaches derives from the combination of controlled experimental designs with their associated normative models, e.g. forced-choice psychophysics experiments with Bayesian observer models. Unfortunately, these tasks do not have much in common with real-world behavior, as they divide behavior into independent trials with discrete responses, often by highly trained participants. In naturalistic tasks, however, behavior is typically continuous and sequential. While highly controlled classical psychophysics tasks allow using normative models to estimate perceptual uncertainty and biases, naturalistic tasks introduce additional cognitive and motor factors such as action variability, intrinsic behavioral costs, and subjective internal models. To account for these factors, I propose to apply inverse normative modeling, i.e. to infer the components of normative models from behavior. In this talk, I will first present recent work that extends Bayesian models of perception to more general cost functions including intrinsic behavioral costs. I will then apply inverse normative modeling to continuous psychophysics. This recently developed experimental approach abandons the rigid trial structure of classical psychophysics and replaces it with a more naturalistic and intuitive continuous tracking task. It produces more temporally fine-grained measurements and allows efficient data collection even with untrained participants. Using Bayesian inverse optimal control, perceptual uncertainty, action variability, behavioral costs, and subjective beliefs about the task dynamics can be estimated from behavior in a tracking task. Finally, I will discuss some limitations of the method and show recent methodological extensions that address these limitations and allow applying inverse optimal control to a wider range of tasks. In summary, these methods open up the possibility of fitting normative models to more naturalistic continuous behavior.

• Speaker: Dominik Straub, TU Darmstadt
• Tuesday 11 February 2025, 15:00-15:45
• Venue: CBL Seminar Room, Engineering Department, 4th floor Baker building.
• Series: Computational Neuroscience; organiser: Daniel Kornai.

Add to your calendar or Include in your list

Certain about uncertainty? Latent representations of VAEs optimized for visual tasks.

Deep Learning methods are increasingly becoming instrumental as modeling tools in Computational Neuroscience, employing optimality principles to build bridges between neural responses and perception or behavior. Deep Generative Models (DGMs) can learn flexible latent variable representations of images while avoiding intractable computations, common in Bayesian inference. However, investigating the properties of inference in Variational Autoencoders (VAEs), a major class of DGMs, reveals severe problems in their uncertainty representations. Here we draw inspiration from classical computer vision to introduce an inductive bias into the VAE by incorporating a global explaining-away latent variable, which remedies defective inference in VAEs. Unlike standard VAEs, the Explaining-Away VAE (EA-VAE) provides uncertainty estimates that align with normative requirements across a wide spectrum of perceptual tasks, including image corruption, interpolation, and out-of-distribution detection. We find that restored inference capabilities are delivered by developing a motif in the inference network (the encoder) which is widespread in biological neural networks: divisive normalization. Our results establish EA-VAEs as reliable tools to perform inference under deep generative models with appropriate estimates of uncertainty.

• Speaker: Josefina Catoni, Universidad Nacional del Litoral
• Friday 14 February 2025, 15:00-16:30
• Venue: CBL Seminar Room, Engineering Department, 4th floor Baker building.
• Series: Computational Neuroscience; organiser: Daniel Kornai.

Add to your calendar or Include in your list

Title to be confirmed

Abstract not available

• Speaker: Sheehan Olver (Imperial College London)
• Thursday 22 May 2025, 15:00-16:00
• Venue: Centre for Mathematical Sciences, MR14.
• Series: Applied and Computational Analysis; organiser: Georg Maierhofer.

Add to your calendar or Include in your list

Synchronization in Navier-Stokes turbulence and it's role in data-driven modeling

n Navier-Stokes (NS) turbulence, large-scale turbulent flows determine small-scale flows; in other words, small-scale flows are synchronized to large-scale flows. In 3D turbulence, previous numerical studies suggest that the critical length separating these two scales is determined by the Kolmogorov length. In this talk, I will introduce our theoretical framework for characterizing synchronization phenomena [1]. Specifically, it provides a computational method for the exponential rate of convergence to the synchronized state, and identifies the critical length based on the NS equations via the “transverse” Lyapunov exponent. I will also discuss the synchronization property of 2D NS turbulence and how it differs from the 3D case [2]. These insights into synchronization and critical length scales are essential for developing machine-learning closure models for turbulence, in particular their stable reproducibility [3]. Finally, I will illustrate how “generalized” synchronization is crucial for predicting chaotic dynamics [4].

[1] M. Inubushi, Y. Saiki, M. U. Kobayashi, and S. Goto, Characterizing small-scale dynamics of Navier-Stokes turbulence with transverse Lyapunov exponents: A data assimilation approach, Phys. Rev. Lett. 131, 254001 (2023).

[2] M. Inubushi and C. P. Caulfield (in preparation).

[3] S. Matsumoto, M. Inubushi, and S. Goto, Stable reproducibility of turbulence dynamics by machine learning, Phys. Rev. Fluids 9, 104601 (2024).

[4] A. Ohkubo and M. Inubushi, Reservoir computing with generalized readout based on generalized synchronization, Sci. Rep. 14, 30918 (2024).

• Speaker: Professor Masanobu Inubushi, Tokyo University of Science
• Friday 14 February 2025, 16:00-17:00
• Venue: MR2.
• Series: Fluid Mechanics (DAMTP); organiser: Professor Grae Worster.

Add to your calendar or Include in your list

A biomass-derived carbon integrated with Zn single-atom catalysts for iodine hosting is designed, integrating the high specific surface area, hierarchical porosity, and nitrogen doping of carbon, along with the excellent chemical confinement and electrocatalytic properties of Zn single atoms. This enabled Zn-I2 batteries to achieve a long lifespan of 80 000 cycles at 10 A g−1 with 93.6% capacity retention.

Abstract

Aqueous zinc iodine (Zn-I2) batteries have attracted attention due to their low cost, environmental compatibility, and high specific capacity. However, their development is hindered by the severe shuttle effect of polyiodides and the slow redox conversion kinetics of the iodine (I2) cathode. Herein, a long-life Zn-I2 battery is developed by anchoring iodine within an edible fungus slag-derived carbon matrix encapsulated with Zn single-atom catalysts (SAZn@CFS). The high N content and microporous structure of SAZn@CFS provide a strong iodine confinement, while the Zn-N4-C sites chemical interact with polyiodides effectively mitigating the iodine dissolution and the polyiodide shuttle effect. Additionally, the uniformly distributed SAZn sites significantly enhance the redox conversion efficiency of I−/I3 −/I5 −/I2, leading to improved capacity. At a high current density of 10 A g−1, the designed Zn-I2 battery delivers an excellent capacity of 147.2 mAh g−1 and a long lifespan of over 80 000 cycles with 93.6% capacity retention. Furthermore, the battery exhibits stable operation for 3500 times even at 50 °C, demonstrating significant advances in iodine reversible storage. This synergistic strategy optimizes composite structure, offering a practical approach to meet the requirements of high-performance Zn-I2 batteries.

Dielectric ceramics with high energy storage performance are crucial for advanced high-power capacitors. Atomic-scale investigations determine that introduction of specific elements (Mg, La, Ca, and Sr) can enhance ferroelectric relaxation behavior by different magnitudes. Disordered polarization distribution and ultrasmall polar nanoscale regions are detected in the high-entropy ceramics after introducing trace amounts of Mg and La. Ultimately, a high recoverable energy density of 10.1 J cm−3 and efficiency of 90% are achieved in a designed high-entropy ceramic.

Abstract

Dielectric ceramics with high energy storage performance are crucial for the development of advanced high-power capacitors. However, achieving ultrahigh recoverable energy storage density and efficiency remains challenging, limiting the progress of leading-edge energy storage applications. In this study, (Bi1/2Na1/2)TiO3 (BNT) is selected as the matrix, and the effects of different A-site elements on domain morphology, lattice polarization, and dielectric and ferroelectric properties are systematically investigated. Mg, La, Ca, and Sr are shown to enhance relaxation behavior by different magnitudes; hence, a high-entropy strategy for designing local polymorphic distortions is proposed. Based on atomic-scale investigations, a series of BNT-based high-entropy compositions are designed by introducing trace amounts of Mg and La to improve the electric breakdown strength and further disrupt the polar nanoscale regions (PNRs). A disordered polarization distribution and ultrasmall PNRs with a minimum size of ≈1 nm are detected in the high-entropy ceramics. Ultimately, a high recoverable energy density of 10.1 J cm−3 and an efficiency of 90% are achieved for (Ca0.2Sr0.2Ba0.2Mg0.05La0.05Bi0.15Na0.15)TiO3. Furthermore, it displays a high-power density of 584 MW cm−3 and an ultrashort discharge time of 27 ns. This work presents an effective approach for designing dielectric energy storage materials with superior comprehensive performance via a high-entropy strategy.

A T g (glass transition temperature) regulation (TR) strategy is developed to effectively release residual strain in the perovskite film through adjusting the ratio of monomeric additives. The resulting film exhibits significantly reduced tensile strain, decreased trap density and superior stability. The optimized perovskite solar cells achieve a high efficiency of 26.15% (certified as 25.59%) and excellent stability.

Abstract

Thermally induced tensile strain that remains in perovskite films after annealing is one of the key reasons for diminishing the performance and operational stability of perovskite solar cells (PSCs). Herein, a glass transition temperature (T g) regulation (TR) strategy is developed by introducing two polymerizable monomers, 2-(N-3-Sulfopropyl-N,N-dimethyl ammonium)ethyl methacrylate (SBMA) and 2-Hydroxyethyl acrylate (HEA), into the perovskite layer. SBMA and HEA undergo in situ polymerization, which regulates the nucleation and crystal growth of the perovskite film. In addition, adjusting the ratio of SBMA and HEA to lower the T g of the resulting polymer effectively releases the strain in the perovskite film. The modified film exhibits significantly reduced tensile strain, decreased trap density and improved stability. As a result, the optimized PSCs achieve a champion power conversion efficiency (PCE) of 26.15% (certified as 25.59%). Furthermore, the encapsulated device demonstrates prominent enhanced operation stability, maintaining 90.3% of its initial efficiency after 500 h of continuous sunlight exposure.

Hybrid MXenes (h-MXenes) are a family of 2D transition metal carbides with amino surface groups that exhibit interesting optical properties. The sensitivity of the organic material under high-energy electron beams complicates conventional TEM analysis. Using cryogenic STEM, we show h-MXenes are stable until a critical dose threshold. Beyond that threshold, the in situ generation of ripplocations is observed.

Abstract

Inorganic–organic hybrid MXenes (h-MXenes) are a family of 2D transition metal carbides and nitrides functionalized with alkylimido and alkylamido surface groups. Using cryogenic and room temperature scanning transmission electron microscopy (STEM) and electron energy-loss spectroscopy (EELS), it is shown that ripplocations, a form of a fundamental defect in 2D and layered structures, are abundant in this family of materials. Furthermore, detailed studies of electron probe sample interactions, focusing on structural deformations caused by the electron beam are presented. The findings indicate that at cryogenic temperatures (≈100 K) and below a specific dose threshold, the structure of h-MXenes remains largely intact. However, exceeding this threshold leads to electron beam-induced deformation through ripplocations. Interestingly, the deformation behavior, required dose, and resultant structure are highly dependent on temperature. At 100 K, it is demonstrated that the electron beam can induce ripplocations in situ with a high degree of precision.

Digital DNA Strand Displacement (DDSD) is developed for advanced molecular classification of processing multiple inputs. By minimizing oligonucleotide requirements, DDSD enables binary and multiclass molecular classification in a simplified manner. It has been successfully applied in infection diagnostics and pathogen identification using blood samples. Computing capability can be further enhanced by cascade DDSD and multiway junction DDSD.

Abstract

DNA-based molecular computing systems for biomarkers have emerged as powerful tools for intelligent diagnostics. However, with the variety of feature biomarkers expanding, current molecular computing systems suffer from the use of a large number of oligonucleotides and limited encoding capability. Here, the study develops an alternative molecular computing approach termed Digital DNA Strand Displacement (DDSD) which recognizes targets and operates target valence through DNA polymerase-based extension and strand release. DDSD significantly reduced the number of used oligonucleotide species, provided robust molecular classifiers. In clinical blood samples, a 96% accuracy rate is achieved with a DDSD-based binary classifier for distinguishing bacterial and viral infections, a 100% accuracy rate is achieved with a multiclass classifier for identifying pathogen types, surpassing existing classifier systems. Moreover, DDSD can be readily expanded. Cascade DDSD is developed, enabling simultaneous computing of up to 14 valence states with a maximum valence of 25. Multiway junction DDSD is implemented to achieve high-valence computing by compact DNA nanostructures rather than split DNA computing units, reducing the potential leakage. The implementation of DDSD enhances the capability of valence-based intelligent molecular diagnostics and multiplexed biomarker detection.

A liposome-based artificial cell is constructed to protect the signal probes and carbon nanotubes are embedded in the artificial cell membranes and employed as artificial channels to enhance intercellular probe transduction and mass transfer. The strategy accelerates cell-cell signal probe transmission and enables effective sensing of let-7a and visualizing let-7a in living colon cancer cells.

Abstract

Monitoring intracellular biomarkers is crucial for clinical disease diagnosis. However, the majority of signaling molecules face difficulties in slow transference across the cell membrane to reach intracellular detection sites, limiting their application in clinical settings. This study proposes an artificial cell-based signal probe transfer-enhanced sensing strategy to effectively detect microRNAs (miRNAs) by embedding carbon nanotubes (CNTs) in artificial cell membranes. The liposome-based artificial cell is constructed to protect the signal probes such as nucleic acids, metal ions, and fluorescent dyes. CNTs are embedded in artificial cell membranes and employed as artificial channels to enhance intercellular signal probe transduction and mass transfer. Through CNTs-mediate cell-cell signal probe transmission, the probes pass through the liposome membrane interface and fuse into target cells, selectively hybridizing with the intracellular target miRNAs, triggering a sensing process and resulting in enhanced fluorescence signal. Furthermore, molecular dynamics simulations are carried out to prove the enhancement of CNTs-mediated cell-cell fusion. This strategy demonstrates excellent analytical performance by quantitatively detecting let-7a miRNA and visualizing it in living colon cancer cells. These findings hold great significance in promoting and accelerating cell-cell signal probe transmission and enabling effective sensing of intracellular biomarkers for diagnostic purposes.

CSAR lecture: The transformative power of empathy for education: reflection and realisation for teaching and learning

Empathy is a facilitator of social, emotional, and cognitive wellbeing and achievement that can actualise teaching, learning and the self. I will present some of my research in empathy over the last three decades from the Institute of Psychiatry King’s College London to the Faculty of Education University of Cambridge, including early childhood empathy, pupil voice, teacher-pupil engagement, creativity in the classroom, and the most recent work using empathy interventions in schools that has shown an increase in empathic awareness, wellbeing and school engagement.

• Speaker: Dr Helen Demetriou, Associate Professor of Psychology and Education, University of Cambridge
• Monday 12 May 2025, 19:30-21:00
• Venue: Location: Wolfson Lecture Theatre, Churchill College, and Zoom.
• Series: Cambridge Society for the Application of Research (CSAR); organiser: John Cook.

Add to your calendar or Include in your list

CSAR lecture: AI in Manufacturing

Opportunities and Challenges for AI in Manufacturing.

All welcome. More details, including a booking link, are here.

• Speaker: Professor Sebastian Pattinson, Institute for Manufacturing, University of Cambridge
• Monday 10 March 2025, 19:30-21:00
• Venue: Location: Wolfson Lecture Theatre, Churchill College, and Zoom.
• Series: Cambridge Society for the Application of Research (CSAR); organiser: John Cook.

Add to your calendar or Include in your list

Equality Saturation in a Real-World Machine Learning Compiler

Join us for a relaxed chat about compilers, while socializing over refreshments. Our social is open to students, academics, professional developers and really anyone interested in compilation. We welcome beginners as well as experts. Our social is an unguided space offered for you to get to know people, try out some new ideas, get feedback on your code, or pair-program on a difficult program. Come with just a paper notebook or bring your laptop to hack on some in-progress patches.

This social is traditionally organized by the LLVM community, but is open to all (potential) compiler enthusiasts. For the first time, the next Compiler Social Talk is part of the department’s Wednesday Seminar.

Equality Saturation in a Real-World Machine Learning Compiler

Machine learning (ML) compilers rely on graph-level transformations to enhance the runtime performance of ML models. However, these program transformations are often driven by manually-tuned compiler heuristics, which are quickly rendered obsolete by new hardware and model architectures. Instead, we propose the use of equality saturation. We replace such heuristics with a more robust global performance model, which accounts for downstream transformations. While this approach still requires a global performance model to evaluate the profitability of transformations,it holds significant promise for increased automation and adaptability. We address challenges in applying equality saturation on real-world ML compute graphs and state-of-the-art hardware, study different cost modeling approaches to deal with fusion and layout optimization, and tackle scalability issues that arise from considering a very wide range of algebraic optimizations. Our implementation builds on and improves the XLA compilation pipeline for CPU and GPU .

This talk is the second one in our compiler social, the first talk will be given by George Constantinides: https://talks.cam.ac.uk/talk/index/225220

Please register if you’re planning to attend: https://grosser.science/compiler-social-2025-02-12/

• Speaker: Alex Zinenko
• Wednesday 12 February 2025, 16:00-20:00
• Venue: Computer Laboratory, William Gates Building, LT1.
• Series: compiler socials; organiser: luisa.cicolini.

Add to your calendar or Include in your list

Feed-Forward Bullet-Time Reconstruction of Dynamic Scenes from Monocular Videos

Hanxue Liang will present his recent work BulletTimer. BulletTimer is the first motion-aware feed-forward model for real-time dynamic scene reconstruction. It takes a monocular video as input and reconstructs a 3D Gaussian Splatting (3DGS) representation at any desired timestamp in a feed-forward manner within just 150 ms. This innovation opens doors for real-time interactive 3D/4D content in areas such as AR/VR, gaming, and virtual cinematography.

• Speaker: Hanxue Liang, University of Cambridge
• Thursday 06 February 2025, 14:00-15:00
• Venue: Webinar (via Zoom online) ID: 87457263100 Passcode: 708172.
• Series: Rainbow Group Seminars; organiser: Yancheng Cai.

Add to your calendar or Include in your list

Energy Environ. Sci., 2024, 17,2610-2620
DOI: 10.1039/D3EE04300B, PaperShanshan Wang, Lin-Yong Xu, Bo Xiao, Mingxia Chen, Meimei Zhang, Wei Gao, Biao Xiao, Alex K.-Y. Jen, Renqiang Yang, Jie Min, Rui Sun
The small molecule donor SD62-doping strategy with an excellent universality is beneficial to fabricating thickness-tolerant and large-scale high-performance polymer solar cells and solar modules.
The content of this RSS Feed (c) The Royal Society of Chemistry

Energy Environ. Sci., 2024, 17,2453-2467
DOI: 10.1039/D4EE00545G, PaperKyra M. K. Yap, William J. Wei, Melanie Rodríguez Pabón, Alex J. King, Justin C. Bui, Lingze Wei, Sang-Won Lee, Adam Z. Weber, Alexis T. Bell, Adam C. Nielander, Thomas F. Jaramillo
Integrated solar fuels devices for CO2 reduction (CO2R) are a promising technology class towards reducing CO2 emissions.
The content of this RSS Feed (c) The Royal Society of Chemistry

Energy Environ. Sci., 2024, 17,2588-2597
DOI: 10.1039/D4EE00433G, PaperSiqi Wang, Yi Wen, Shulin Bai, Zhe Zhao, Yichen Li, Xiang Gao, Qian Cao, Cheng Chang, Li-Dong Zhao
The thermoelectric properties of n-type PbSe are enhanced by optimizing the power factor through crystal growth and slight-tuning vacancy and interstitial defects.
The content of this RSS Feed (c) The Royal Society of Chemistry

Energy Environ. Sci., 2024, 17,2651-2661
DOI: 10.1039/D3EE04253G, PaperGui Li, Jian Wang, Yue He, Shuyan Xu, Shaoke Fu, Chuncai Shan, Huiyuan Wu, Shanshan An, Kaixian Li, Wen Li, Ping Wang, Chenguo Hu
A new type of sliding-TENG with asymmetric electrode design achieves high output performance and ultra-stability by the balance of electrostatic shielding and charge accumulation for providing power supply to some electrical devices.
The content of this RSS Feed (c) The Royal Society of Chemistry

“Innate immune sensing of viral nucleic acids"

This Cambridge Immunology and Medicine Seminar will take place on Thursday 6 February 2025, starting at 4:00pm, in the Ground Floor Lecture Theatre, Jeffrey Cheah Biomedical Centre (JCBC)

Title: “Innate immune sensing of viral nucleic acids”

Speaker: Dr. Brian Ferguson, is Associate Professor of Immunology, Department of Pathology, University of Cambridge.

Host: Margaret Stanley, Emeritus Professor, Cambridge

Refreshments will be available following the Seminar.

• Speaker: Dr. Brian Ferguson, Associate Professor, Department of Pathology
• Thursday 06 February 2025, 16:00-17:00
• Venue: Lecture Theatre, Jeffrey Cheah Biomedical Centre, Cambridge Biomedical Campus.
• Series: Cambridge Immunology Network Seminar Series; organiser: Ruth Paton.

Add to your calendar or Include in your list