About

Ani Liu is the Carrafiell Assistant Professor of Fine Arts at the Weitzman School of Design, University of Pennsylvania, having joined the faculty in fall 2021. Her artistic practice explores the effects of emerging technologies on our lives and communities, using science and technology to create emotional experiences. Her recent projects include works on view at the Museum of Arts and Design in New York and a course she teaches on research in art and design. Liu's work has increasingly centered on motherhood, examining its physiological and emotional transformations through a lens of research and artistic expression. She has created sculptures and installations that document and reflect on the labor of care, such as 'Untitled (Labor of Love)' and 'Untitled (pumping),' which highlight the invisible work involved in caregiving and breastfeeding. Her recent focus also includes microplastics, investigating their presence in the human body, particularly in breast milk, and contemplating the systemic issues related to plastic consumption. Liu's work often involves collaboration and conversation, including engaging with trans and nonbinary individuals about gender, fertility, and parenthood, as exemplified by her piece 'Untitled (pregnancy menswear).' She emphasizes the importance of research in her artistic practice, exploring how knowledge is produced and how drawing and observation serve as forms of research and thinking. Her approach integrates interdisciplinary inquiry, combining art, science, and social issues to challenge and expand perceptions of gender, care, and environmental impact.

Research topics

• Artificial Intelligence
• Computer Science
• Distributed computing
• Machine Learning
• Political Science
• Materials science
• Medicine
• Family medicine
• Chemical physics
• Composite material
• Library science
• Crystallography
• Thermodynamics

Selected publications

• Contrastive learning in tunable dynamical systems

  arXiv (Cornell University) · 2026-03-27

  preprintOpen access

  We generalize the theory of supervised contrastive learning, previously applied to physical systems at equilibrium or steady state, to systems following any dynamics described by coupled ordinary differential equations. We show that if physical dynamics break time reversal symmetry, gradient descent on a cost function embodying the desired behavior cannot be achieved with a scalable process, even in principle. We therefore introduce Probably Approximately Right (PAR) learning processes, composed of a local contrastive learning rule and a scalable supervision protocol. We show that approximate, local supervision with forward propagation of the error signal can be used to successfully train several tunable models of physical dynamics inspired by examples in biological and machine learning.

  PublisherDOI

• Unsupervised and probabilistic learning with Contrastive Local Learning Networks: The Restricted Kirchhoff Machine

  Proceedings of the National Academy of Sciences · 2026-05-21

  preprintOpen accessSenior author

  Autonomous physical learning systems modify their internal parameters and solve computational tasks without relying on external computation. Compared to traditional computers, they enjoy distributed and energy-efficient learning due to their physical dynamics. In this paper, we introduce a self-learning resistor network, the Restricted Kirchhoff Machine, capable of solving unsupervised learning tasks akin to the Restricted Boltzmann Machine algorithm. The circuit relies on existing technology based on Contrastive Local Learning Networks, in which two identical networks compare different physical states to implement a contrastive local learning rule. We simulate the training of the machine on a dataset of handwritten digits, providing a proof of concept of its learning capabilities. Finally, we compare the scaling behavior of time, power, and energy per operation as the number of nodes increases to that of a Restricted Boltzmann Machine implemented on central processing unit (CPU) and graphics processing unit (GPU) platforms.

  PublisherDOI

• Contrastive learning in tunable dynamical systems

  arXiv (Cornell University) · 2026-03-27

  articleOpen access

  We generalize the theory of supervised contrastive learning, previously applied to physical systems at equilibrium or steady state, to systems following any dynamics described by coupled ordinary differential equations. We show that if physical dynamics break time reversal symmetry, gradient descent on a cost function embodying the desired behavior cannot be achieved with a scalable process, even in principle. We therefore introduce Probably Approximately Right (PAR) learning processes, composed of a local contrastive learning rule and a scalable supervision protocol. We show that approximate, local supervision with forward propagation of the error signal can be used to successfully train several tunable models of physical dynamics inspired by examples in biological and machine learning.

  PublisherOA PDF

• Training of physical neural networks

  Nature · 2025-09-03 · 45 citations

  reviewOpen access
  PublisherOA PDFDOI

• Roadmap on machine learning glassy dynamics

  Nature Reviews Physics · 2025-01-06 · 29 citations

  reviewOpen access
  PublisherOA PDFDOI

• Mechanosensitive Remodeling Sustains Rigidity Homeostasis in Actin Cortex Models

  bioRxiv (Cold Spring Harbor Laboratory) · 2025-10-06 · 2 citations

  preprintOpen accessSenior authorCorresponding

  The actin cortex is a dynamic biopolymer network whose mechanical rigidity, while relying critically on tensioned filaments, is robustly sustained amid constant architectural changes through the assembly and disassembly of filaments and crosslinkers. Yet the role of such remodeling processes in rigidity homeostasis remains essentially unexplored in computational models. As a result, we still lack proper understanding of the biological rationale for remodeling, which is energetically expensive, or of the microscopic mechanisms through which collective rigidity is maintained. To address this, we develop two complementary elastic network models in which rigidity homeostasis with complete turnover emerges as a result of mechanosensitive dynamics of filaments (edges) and crosslinkers (nodes), respectively. Both models require the following minimal ingredients: (1) preferential disassembly of edges or nodes under small tension or force, (2) a small but nonzero rate of random disassembly, and (3) energy injection upon assembly. Our models are robust to variations in random disassembly rates and can recover from drastic structural disruption. Remarkably, nodes and edges undergo diffusion even while elastic moduli and structural correlations reach steady states, showing that the models display representational drift similar to that found in neuronal activities and physical learning circuits. We propose that the cortex is an example of “tunable matter,” i.e ., its mechanosensitive remodeling dynamics tune its edges and nodes so that the cortex as a whole can maintain robust but flexible rigidity in fluctuating mechanical environments, creating survival advantages that justify its energy consumption.

  PublisherOA PDFDOI

• Cooperative Function with Thermal Fluctuations in Mechanical Networks

  ArXiv.org · 2025-09-24

  preprintOpen access

  Elastic networks can be tuned to exhibit complex mechanical responses and have been extensively used to study protein allosteric functionality, where a localized strain regulates the conformation at a distant site. We show that cooperative binding, where two sites each enhance the other's ability to function, can be trained via a symmetric application of the training previously employed for creating network allostery. We identify a crossover temperature above which cooperative functionality breaks down due to thermal fluctuations. We develop a modified training protocol to increase this crossover temperature, enabling function to remain robust at biologically relevant temperatures.

  PublisherOA PDFDOI

• Microscopic Imprints of Learned Solutions in Tunable Networks

  Physical Review X · 2025-07-15 · 1 citations

  articleOpen accessSenior author

  In physical networks trained using supervised learning, physical parameters are adjusted to produce desired responses to inputs. An example is an electrical contrastive local learning network of nodes connected by edges that adjust their conductances during training. When an edge conductance changes, it upsets the current balance of every node. In response, physics adjusts the node voltages to minimize the dissipated power. Learning in these systems is therefore a coupled double-optimization process, in which the network descends both a cost landscape in the high-dimensional space of edge conductances and a physical landscape—the power dissipation—in the high-dimensional space of node voltages. Because of this coupling, the physical landscape of a trained network contains information about the learned task. Here, we derive a structure-function relation for trained tunable networks and demonstrate that all the physical information relevant to the trained input-output relation can be captured by a tuning susceptibility, an experimentally measurable quantity. We supplement our theoretical results with simulations to show that the tuning susceptibility is correlated with functional importance and that we can extract physical insight into how the system performs the task from the conductances of highly susceptible edges. Our analysis is general and can be applied directly to mechanical networks, such as networks trained for protein-inspired function such as allostery.

  PublisherDOI

• Epithelial convergent extension as a tuning process

  bioRxiv (Cold Spring Harbor Laboratory) · 2025-11-07

  preprintOpen access

  Self-tuning—the ability of disordered systems to develop desired collective behaviors by tuning internal couplings in response to feedback—has recently emerged as a powerful framework for understanding adaptation in amorphous solids, mechanical metamaterials, and electrical networks. These systems can learn desired responses, encode memory, and robustly reorganize under repeated stimuli, much like artificial neural networks but without requiring processors to adjust their weights. Here, we extend this paradigm to morphogenesis and show that the epithelium can be viewed as tunable matter and that epithelial convergent extension (CE) can be understood as a self-tuning process. Using a vertex model with active interfacial tensions, we systematically compare distinct tension-update strategies, including externally imposed shear, global gradient descent optimization, and decentralized local feedback rules. We find that while all methods can generate tissue elongation, only local orientation- and length-sensitive rules reproduce key experimental features of CE, such as supracellular actomyosin pattern formation, cell shape changes, and junctional alignment. In contrast, global optimization produces homogeneous tension patterns and mechanically fragile states. By interpreting CE through the lens of tuning, our framework bridges the physics of tunable matter with developmental biology, revealing how simple, local rules enable tissues to efficiently orchestrate complex morphogenetic outcomes through decentralized mechanical adaptation.

  PublisherOA PDFDOI

• Rigidity of epithelial tissues as a double optimization problem

  Physical Review Research · 2025-02-12 · 12 citations

  articleOpen access

  How do cells tune emergent properties at the scale of tissues? One class of such emergent behaviors are rigidity transitions, in which a tissue changes from a solidlike to a fluidlike state or vice versa. Here we introduce a way for a tissue described by a vertex model to tune its rigidity by using “tunable degrees of freedom.” We use the vertex model elastic energy as a cost function and the cell stiffnesses, target shapes, and target areas as different sets of degrees of freedom describing cell-cell interactions that can be tuned to minimize the cost function. We show that the rigidity transition is unaffected when cell stiffnesses are treated as tunable degrees of freedom. When preferred shapes or areas are treated as tunable degrees of freedom, however, induced spatial correlations in target cell shapes or areas shift the rigidity transition. These observations suggest that tissues can coordinate changes in cell-scale properties, treated here as tunable degrees of freedom, to achieve desired tissue-scale behaviors.

  PublisherDOI

Recent grants

• Self-assembly of Charged Biopolymers in Solution

  NSF · $519k · 2001–2006

• Theoretical Studies of Tunable Networks

  NSF · $670k · 2021–2026

• Self-assembly and motility far from equilibrium

  NSF · $545k · 2011–2016

• Statistical Physics of Disordered and Driven Systems

  NSF · $300k · 2006–2010

• Self-assembly of Charged Biopolymers in Solution

  NSF · $109k · 2005–2008

Frequent coauthors

• Sidney R. Nagel

  University of Chicago

  105 shared

• D. J. Durian

  79 shared

• Menachem Stern

  University of Pennsylvania

  54 shared

• Sean A. Ridout

  49 shared

• Carl P. Goodrich

  Institute of Science and Technology Austria

  43 shared

• Sam Dillavou

  California University of Pennsylvania

  42 shared

• Samuel S. Schoenholz

  Google (United States)

  35 shared

• Benjamin D. Beyer

  University of Pennsylvania

  32 shared

Education

• Ph. D., Physics

  Cornell University

  1989

• B. A., Physics

  University of California Berkeley

  1984