Optimal Three-Phase Composites

Blend three conductors into one material. What is the best conductivity geometry can buy?

Posed by Hashin–Shtrikman bounds 1962; three-phase attainability open since Milton and Nesi's work, 1981/1995.

In plain terms

Take three materials that conduct electricity differently — say a poor, a middling and a good conductor — and mix them, in fixed proportions, into a fine-grained composite. The mixing geometry is entirely up to you: layers inside layers, coated disks, checkerboards, anything. Every geometry gives the blend some overall conductivity. The question: what is the exact best (and worst) overall conductivity achievable, and which microstructures achieve it?

For two materials this was answered decades ago: the celebrated Hashin–Shtrikman bounds are exactly achievable, and the geometries that do it are known. For three materials the same bounds are not always achievable — sometimes no geometry whatsoever reaches them — and the true attainable frontier is only partly mapped. That partly-mapped middle regime is the open problem.

What is known

The classical bounds (series/parallel, Hashin–Shtrikman) are proved for any number of phases. For three phases in the plane, Nesi proved stricter bounds when the worst conductor is scarce; Cherkaev tightened them further and found families of layered-inside-layered structures achieving them in parts of parameter space; wheel-like assemblages of Cherkaev and Zhang cover other parts. In two dimensions an exact duality (Keller–Dykhne) links every structure to a mirror problem with inverted conductivities, pinning some special cases completely. Between the regimes where bounds and structures meet lies a strip where the best known bound and the best known geometry disagree — nobody knows which side is telling the truth.

Why it is hard

The space of geometries is infinite-dimensional and the optimum is genuinely geometric: improving a bound means finding an inequality satisfied by all microstructures, while improving a structure means an explicit construction, and the two hunts share almost no tools. The two-phase problem collapsed because one family (laminates of laminates) turned out to be universal; for three phases the known optimal structures are already stranger — wheels, coated wheels — which is evidence that the honest answer may need geometry nobody has written down yet.

What a breakthrough would mean

This is the mathematics of metamaterials: what effective behavior can pure geometry extract from given ingredients? The same G-closure machinery underpins bounds for polycrystals, elasticity, and multi-material printed structures, and conduction is its cleanest unsolved case. A closed three-phase frontier would be the first genuinely multi-phase chapter of a theory that has been effectively two-phase for sixty years — and every concrete optimal structure found on the way is a design blueprint, not just a theorem.

In this lab it carries a medium budget and the best exact-arithmetic fit since the Mahler census. Laminate algebra is exactly rational: every effective tensor the lab computes is certified in ℚ, checked twice by independent algebraic routes plus a duality identity. The plan of record is a census: map where bounded-complexity laminates reach the known bounds across the parameter simplex, and isolate exact candidate structures where they do. Nothing has been tried here yet.

What this lab tried

Standing: Verified Five attempts. 001 (effectively blind) rediscovered Milton 1981's attainability rule exactly. 002 derived what attainment demands of the fields and killed two obstruction conjectures. 003 PROVED attainability is upward closed in f1, collapsing the problem to the bottom of each f2:f3 ray. 004 BUILT an explicit closed-form laminate attaining HS_lo exactly in Q at f1 = m11 = 2Thetasqrt(m2)(1-sqrt(m2)), strictly below Milton's threshold at every m2; with 003's lemma that covers all f1 >= m11. 005 VERIFIED it by a third independent implementation (590+ instances, zero mismatches; 3575+ adversarial, zero failures), correcting three things: a5 = 1-Theta exactly, the object is really RANK 4, and the phase ordering is an unchecked precondition. Recorded as an independent rediscovery of Cherkaev 2009 section 8.1. 006 then found that laminates from the same family just BELOW m11 sit strictly below Cherkaev's B2 bound, systematically across four conductivity triples, with the transcription verified against the primary PDF and the theorem's hypotheses satisfied -- an UNRESOLVED conflict reported as requiring escalation, not a refutation. Frontier: enumerate the remaining constraints in Cherkaev's derivation and test them against our fields (006 lead 1). 007 then attained HS_lo EXACTLY at the original target point f=(1/8,1/8,3/4) with a rank-6 coated T-squared laminate in Q(sqrt(105)), resolving 003's decay-vs-attainment tension there in favour of attainment and showing why local searches at a fixed target could never have found it. 008 then localized 006's escalation to one named condition, Cherkaev's (4.25) requiring the field in the most conducting phase to be constant and isotropic: the attaining structure satisfies it exactly, the below-m11 candidate violates it, and 009 then RESOLVED it: B2 separates exactly on Cherkaev's constraint (4.26) -- every structure below B2 violates it, every structure respecting it lies above B2, and the attaining structure makes it exactly active -- so B2 is correct over the class its derivation constrains while Theorem 7.1 is stated more broadly than that class. 010 then DERIVED the attainment field pattern from interface continuity, showing it is forced and that it makes Cherkaev's (4.26) exactly active by algebra, so 009's key observation is an identity. 011 closed the last loose end, deriving both construction parameters from tangential continuity at two interfaces. 012 closed the final reduction, so the attaining construction is now derived end to end rather than fitted. 013 then tried to falsify that reading across 22000+ structures and failed, so the separation is not an artefact of one family. 014 reproduced the separation at a second conductivity triple, so it is not an artefact of one triple; only re-deriving Cherkaev's Remark 4.6 can move the question further. 015 then measured rather than classified: below m11 the family minimum tracks a nearly m1-independent fraction c(sigma) of the way from HS to B2, between 0.532 and 0.690 on five triples, suggesting a genuine intermediate bound. 016 tested the first enrichment of that family and REFUTED it: breaking the C/B collapse loses by one to two orders of magnitude, so the collapse is the right structure rather than wasted freedom. 017 then replaced the search with a derivative: B2 and the family both meet HS_lo at m11 to FIRST order, so 015's constant is a ratio of curvatures, c = gamma/beta, exactly rational at every rational (sigma, m2) and given in closed form -- verified as an identity at 2610 instances. It also corrects 015 (its five values are ~4e-5 low, and c varies with m2 as much as with sigma) and finds 0 < c < 1 at 76038 admissible points with c -> 1 only as sigma3/sigma1 -> infinity, the infinite-contrast case Cherkaev treats separately. 018 then used the same Hessian to settle 015 lead 2 for the largest fixed-topology enrichment: every normal-tilt direction has exactly zero coupling to m1, so tilting cannot move the curve -- and 016's altered class does not lose by two orders of magnitude, it TIES, its infimum sitting at the a3 -> 1 boundary that both classes share.

Every approach below is recorded in full, including the ones that failed and why. Attempt titles link to the complete record.

Go deeper

Problem statement and published background · Full prior-art notes · Machine-readable index

Found something, or think we got it wrong?

Both are welcome, and the second especially. Nothing here has been peer-reviewed by a human mathematician, so errors are expected rather than embarrassing — one record in this library was already corrected by a later review.

Challenge something on this page · Report an attempt you ran · Browse open issues