Memory Elements: A Paradigm Shift in Lagrangian Modeling of Electrical Circuits Dimitri Jeltsema Delft Institute of Applied Mathematics Delft University of Technology The Netherlands Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 1 Classical Circuit Theory Classical (linear) circuit theory defines three basic elements: I Resistor R I Inductor L I Capacitor C each described by a relationship between two of the four basic variables I Voltage V I Current I I Flux φ I Charge q Rt Rt where φ(t ) = −∞ V (τ )dτ , and q (t ) = −∞ I (τ )dτ. Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 2 Are We Missing Something? Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 3 The Memristor ‘Missing’ relationship between φ and q defines the memristor.* I Charge-controlled: φ = φ̂(q ) or by using φ̇ = V and q̇ = I V = RM (q )I where RM (q ) := dφ̂ dq (q ) is the incremental memristance. I The memory-effect stems from the fact that Z t q (t ) = I (τ )dτ, −∞ i.e., a memristor bookkeeps the current flowing through it! *L.O. Chua. Memristor—the missing circuit element. IEEE Trans. Circ. Theory, CT–18(2):507–519, September 1971. Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 4 Mechanical Memristor Perhaps more transparent in the mechanical domain. By analogy: = Rt I Displacement: x (t ) = I Momentum: p (t ) −∞ F (τ )dτ Rt −∞ v (τ )dτ Example: tapered dashpot. Displacement dependent ⇒ p = p̂(x ) F F x I Diff. wrt time yields F = RM (x )v with RM (x ) := Delft Institute of Applied Mathematics dp̂ dx (x ). MTNS 2012 — Melbourne Australia 5 Mechanical Memristor Blueprint for memristors are pinched hysteresis loops: I Sinusoidal excitation: v (t ) = sin(t ) I Two-valued everywhere (except at origin) Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 6 Mechanical Memristor Underlying single-valued characteristic: p(t ) = 13 x 3 (t ) h i2 Rt ⇒ F (t ) = x (0) + 0 v (τ )dτ v (t ) | {z } RM (x (t )) Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 7 Hodgkin-Huxley Model In 1952, Nobel laureates Hodgkin and Huxley suggested a circuit model (left) that represents represent the biophysical characteristic of cell membranes. However, they (mis-)interpreted the conductances as time-varying. It is shown∗ that GNa and GK actually correspond to memristors (right). *L.O. Chua and S.M. Kang. Memristive devices and systems, Proc. IEEE, 64:209-223, 1976. Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 8 Improved Josephson Junction Circuit Model More rigorous analysis has revealed presence of additional small current I = g cos(ko φM )V , for some constants g , ko ⇒ flux-controlled memristor: qM = q̂M (φM ) = g ko sin(ko φM ), with q̇M = I and φ̇M = V . ∗ T.P. Orlando and K.A. Delin, Foundation of Superconductivity. Reading, MA: Addison-Wesley, 1991. Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 9 HP’s (Hewlett-Packard) Memristor Meanwhile, researchers at HP are trying to create alternative for flash memory using a two-layer semiconductor constructed from layers of titanium oxide. Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 10 HP’s (Hewlett-Packard) Memristor *D.B. Strukov, G.S. Snider, D.R. Stewart, and R.S. Williams. The missing memristor found. Nature, 453:80–83, May 2008. Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 11 Other Interesting Applications I Non-volatile nano memory I Crossbar latches as transistor replacements I Analog computing I Programmable logic and signal processing I Learning circuits I Models for biological phenomena I Quantum computing I ??? Check June 2012 special issue on Memristor Technology in Proc IEEE. Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 12 But the story does not end... It actually just started... Rt Let us also consider: σ(t ) = −∞ q (τ )dτ, I Define ρ(t ) = Rt −∞ φ(τ )dτ ρ = ρ̂(q ) . Then, differentiation wrt time yields φ = LM (q )I ⇒ Meminductor where LM (q ) := I Define dρ̂ dq (q ) is the incremental meminductance. σ = σ̂(φ) . Then, differentiation wrt time yields q = CM (φ)V where CM (φ) := dσ̂ dφ ⇒ Memcapacitor (φ) is the incremental memcapacitance. *M. Di Ventra et al. (2009). Circuit elements with memory: memristors, memcapacitors and meminductors. Proc. of the IEEE, 97(10), 1717-1724. Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 13 Family of Memory Elements (Mem-elements) Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 14 Btw: memcapacitor? Or is it FLUX Capacitor? Remember this guy... Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 15 Classical Circuit-Theory at its Limits I Straightforward application of conventional Lagrangian modeling to circuits with memory elements may yield the wrong equations. I We provide a novel variational method to overcome these problems using a different set of configuration coordinates I The new Lagrangian equations do not correspond to KVL (Kirchhoff voltage law) or KCL (Kirchhoff current law) anymore, but represent conservation of flux and charge, i.e., iKVL and iKCL. I Furthermore, we show that our configuration coordinates are electrical analogies of a mechanical quantity called “absement”. Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 16 Classical Lagrangian Approach I If the inductor is current-controlled, we need to consider loop analysis and define q̇ Z φ̂(I )dI − L(q , q̇ ) = 0 1 2C q2, which upon substitution into d ∂L dt ∂ q̇ − ∂L q = 0, yields φ̂0 (q̇ )q̈ + = 0. ∂q C I Or a node analysis in case of a flux-controlled inductor, i.e., 1 ∗ 2 Z L (φ, φ̇) = C φ̇ − 2 φ Î (φ0 )dφ0 , 0 which together with d ∂L∗ dt ∂ φ̇ − ∂L∗ = 0, yields C φ̈ + Î (φ) = 0. ∂φ Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 17 Classical Lagrangian Approach Now, replace conventional inductor by a meminductor... The Lagrangian becomes (q̇ = I) 1 1 2 2C L(q , q̇ ) = LM (q )q̇ 2 − q2, generating the equation of motion LM (q )q̈ + q 1 0 LM (q )q̇ 2 + = 0. 2 C However, if we just apply KVL, we obtain (recall φ = LM (q )I) φ̇ + q C = LM (q )İ + L0M (q )I 2 + q C = 0. The Lagrangian framework generates an erroneous factor 12 ! Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 18 Two Problems (1) Path-dependence: magnetic (co-)energy 1 L 2 M (q )q̇ 2 depends on q. (2) Self-adjointness: Necessary condition to find L such that d ∂L dt ∂ ẋ i − ∂L ≡ Aij (x , ẋ )ẍ j + Bi (x , ẋ ) = 0, ∂x i is given by the integrability conditions (see Santilli (1978)) Aij = Aji , ∂ Aik ∂ Ajk = , j ∂ ẋ ∂ ẋ i ∂ Bj 1 ∂ ∂ Bi ∂ Bj k ∂ Bi − i = − i ẋ , ∂x j ∂x 2 ∂ x k ∂ ẋ j ∂ ẋ ∂ Bi ∂ Bj ∂ Aij + i = 2 k ẋ k . ∂ ẋ j ∂ ẋ ∂x Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 19 Not Self-Adjoint... I It is directly verified that, with A(q ) = LM (q ) and B (q , q̇ ) = L0M (q )q̇ 2 + q /C ⇒ not self-adjoint ⇒ no Lagrangian formulation as such. I Possible solution: introduce Rayleigh dissipation function 1 D(q , q̇ ) = L0M (q )q̇ 3 , 6 so that LM (q )q̈ + | 1 0 q ∂D LM (q )q̇ 2 + =− . 2{z C} ∂ q̇ d ∂L − ∂L dt ∂ q̇ ∂q However, variational principle is lost :-( Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 20 Conservation of Flux and Charge The conventional Lagrangian formalism essentially codes the KVL and KCL in terms of energy storage via the Lagrangian. Lets consider iKVL and iKCL X φi (t ) = j Vi (τ )dτ, −∞ i X t Z φi (t ) = 0, Z qj (t ) = 0, t qj (t ) = Ij (τ )dτ, −∞ respectively (flux and charge can neither be created nor be destroyed). Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 21 Memory State Functions Recall that fundamental relationship of a meminductor reads ρ = ρ̂(q ) For that reason, let us instead of the magnetic (co-)energy define a state function of the form ∗ Z T̄ (q ) := ρ̂(q )dq Furthermore, since a linear memcapacitor reduces to a conventional linear capacitor, i.e., σ = C φ, we may associate it with a function of the form Ū (σ) := 1 2C σ2 Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 22 An Alternative Variational Principle Then select, instead of a loop charge q, the integrated loop charge σ , with σ̇ = q, as the configuration variable, and define the Lagrangian σ̇ Z L̄(σ, σ̇) = ρ̂(q )dq − 0 1 2C σ2. Indeed, invoking Hamilton’s principle by considering variations in terms of σ , we get the Lagrangian type of equation d ∂ L̄ dt ∂ σ̇ − ∂ L̄ = 0, ∂σ which, in turn, generates ρ̂0 (σ̇)σ̈ + σ/C = 0 (where ρ̂0 (σ̇) =: LM (q )). Differentiating the latter with respect to time yields ... ρ̂0 (σ̇) σ + ρ̂00 (σ̇)σ̈ 2 +σ̇/C = 0 | {z } ≡ φ̇ + q C . φ̇ Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 23 Integrated Charge = Electrical “Absement” Classical Analogy: Charge Displacement In mechanics, displacement x and its various derivatives define an ordered hierarchy of meaningful concepts: I 1st derivative of displacement is velocity I 2nd derivative is acceleration I 3rd derivative is jerk I 4th derivative is jounce I etc. Recently, also -1st, -2nd, -3rd, etc. derivatives introduced to model flow-based musical instruments*. *S. Mann et al. (2006). Hydraulophone design considerations: absement, displacement, and velocity-sensitive music keyboard in which each key is a water jet. In Proc. 14th annual ACM int. conf. on Multimedia, Santa Barbara, 519-528. Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 24 Integrated Charge = Electrical “Absement” I Absement = contraction of ‘absence’ and ‘displacement’ I In SI units measured in meter times seconds [ms] I One meter-second corresponds to being absent one meter from an origin or other reference point for a duration of one second I integrated charge absement I absement of a capacitor is a measure for the amount of charge that is needed in a particular time interval to charge the capacitor to a certain level I For a memcapacitor absement is as a measure of how much charge and time is needed to write its memory. Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 25 Concluding Remarks I Memory elements become increasingly important. I Memory elements mimic behavior of synapse. I Create biologically-inspired electronic circuits. I Various interesting applications. I HP announced that memristive random access memory may reach the market as early as 2013. I Teaching memory elements is a sine qua non for the training of the next generation of scientists and engineers. I Classical circuit- and systems theory needs to be modified and/or extended. I Absement a natural and important variable in this context. Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 26 THANKS! And never forget that... “The only things you are going to regret in this life are the risks you didn’t take.” Grumpy Old Men (1993) Delft Institute of Applied Mathematics MTNS 2012 — Melbourne Australia 27