Material-Based Intelligence: Self-organizing, Autonomous and Adaptive Cognition Embodied in Physical Substrates

TL;DR

Material-Based Intelligence (MBI) addresses the gap where materials autonomously sense, learn, and adapt through embedded physics rather than external computation. The paper defines architectural foundations and functional manifestations, proposing three strategies—multistability, dissipative adaptation, and criticality—to realize emergent intelligence in soft and active matter. It introduces an experimental roadmap including an MBI Testing Arena and curated databases, with metrics to quantify information integration, memory fidelity, adaptation, autonomy, and robustness. The work argues that advancing MBI could yield autonomous soft robots, adaptive materials, and new insights into cognition, moving beyond biomimicry toward fully synthetic, self-evolving systems.

Abstract

The design of intelligent materials often draws parallels with the complex adaptive behaviors of biological organisms, where robust functionality stems from sophisticated hierarchical organization and emergent long-distance coordination among a myriad local components. Current synthetic materials, despite integrating advanced sensors and actuators, predominantly demonstrate only simple, pre-programmed stimulus-response functionalities, falling short of robustly autonomous intelligent behavior. These systems typically execute tasks determined by rigid design or external control, fundamentally lacking the intricate internal feedback loops, dynamic adaptation, self-generated learning, and genuine self-determination characteristic of biological agents. This perspective proposes a fundamentally different approach focusing on architectures where material-based intelligence is not pre-designed, but arises spontaneously from self-organization harnessing far-from-equilibrium dynamics. This work explores interdisciplinary concepts from material physics, chemistry, biology, and computation, identifying concrete pathways toward developing materials that not only react, but actively perceive, adapt, learn, self-correct, and potentially self-construct, moving beyond biomimicry to cultivate fully synthetic, self-evolving systems without external control. This framework outlines the fundamental requirements for, and constraints upon, future architectures where complex, goal-directed functionalities emerge synergistically from integrated local processes, distinguishing material-based intelligence from traditional hardware-software divisions. This demands that concepts of high-level goals and robust, replicable memory mechanisms are encoded and enacted through the material's inherent dynamics, inherently blurring the distinction between system output and process.

Figures (4)

• Figure 1: The Conventional Control Architecture and its Separation of Physical and Computational Domains. Conventional control architecture of an embodied agent, i.e. a controlled system that performs some actions in the physical world. Such an agent requires at least a twofold separation or encapsulation of physical dynamics and control. First, the computing machine including its communication channels with sensors and actuators establishes an interface between abstract and portable algorithms and the computing hardware. The according flow of information is depicted by the vertical and the topmost arrows. The reduced thickness of the vertical arrows visualizes the fact that the sensory bandwidth is still a bottleneck in control. The centred flow of information (decorated with the term "engineered design") represents the part of the control that is facilitated by proper planning and design. Thirdly, there is what we call noise in this context, i.e., the information flow resulting from external influences (which are not necessarily known in all details.) Those parts of the dynamics that are related to the (known) engineered design are modelled, whereas noise stands for the unmodelled parts. Note the importance of a de-physicalization of sensory input and the re-physicalization of the result of computations with control purpose by actuators. This de- and re-physicalization is a characteristic feature of systems with a clear separation between hard- and software and may be obsolete in material-based intelligence (Figure adapted from fuchslin_morphological_2013).
• Figure 2: The Landscape of Intelligent Systems and the vast land for Material-Based Intelligence. This schematic maps different forms of intelligence onto a design space defined by two key axes: the degree of hardware-software separation (horizontal) and the breadth of problem-solving capability (vertical). Distinct lines of development traverse this landscape. The mathematical/formal Line (purple) illustrates increasing computational power within a paradigm of high abstraction and hardware independence. The (bio-)evolutionary Line (green) shows how natural systems achieved sophisticated cognitive capabilities (e.g., cells, immune systems, brains) while maintaining a deep integration of information processing with their physical bodies. This line defines a trajectory deep into the high-integration, high-functionality quadrant. The technological Line (red) and a branch, the optimization line (magenta) charts the development from embodied tools to electro-mechanical systems, before making a significant leap to abstract digital computation, thereby leaving a conceptual "gap" in the development of highly capable yet physically integrated systems. The jump from analogue controllers to machines capable of various degrees of formal computation is wide and consequently, the line connecting them is dashed. Material-Based Intelligence (MBI) is positioned to fill this gap, aiming to engineer systems that exhibit complex, autonomous functions directly through their physical structure and dynamics. This paradigm sits at the intersection of biological inspiration and a new technological frontier, leveraging the principles of embodiment over pure abstraction. There are outliers not covered by lines in this diagram. As a prominent example, we noted protein synthesis. Although it relies deeply on the details of physics and chemistry, the process can produce a combinatorially rich number of outcomes in a programmable manner.
• Figure 3: The Material-Based Intelligence (MBI): From Physical Ingredients to Emergent Function. This diagram provides a view of the MBI paradigm, organised into foundational elements, core processes, and emergent outcomes, illustrating a bottom-up flow from physical substrates to functional intelligence.
• Figure 4: The Dynamical Systems Approach to Material-Based Intelligence. (Left) Conceptual attractor landscape for a dynamic MBI system with two parameters p$_1$ and p$_2$. The system's state (representing, for example, gaits in a robot or reaction product concentrations) naturally evolves towards an attractor, for devices with a purposeful function a stable state or a limit cycle (e.g., I$_1$, I$_2$). External control input or environmental cues. Assuming a tuning parameter $\alpha$, one can reshape this landscape, altering the basins of attraction (dashed lines) and causing the system to transition between these functional modes (double-headed arrows). This mechanism underpins morphological control, where the detailed physical execution of a pattern is handled autonomously by the system's dynamics. (Right) A Finite State Machine (FSM) can be emulated by chaining multiple such dynamical systems. The final output state (characterised e.g. by p$_1^f$ and p$_2^f$) of one dissipative dynamical system (DS$_1$), once settled into an attractor from its initial condition I$_1$, serves as the input (e.g., I$_2$) or as a parameter to tune the attractor landscape of the next dynamical system (DS$_2$). This physical propagation of information can extend to sequences or loops of systems, embodying a discrete computational flow through a continuous physical medium.