diff --git a/src/assemblers/abstract.jl b/src/assemblers/abstract.jl
index 865e374..ac59836 100644
--- a/src/assemblers/abstract.jl
+++ b/src/assemblers/abstract.jl
@@ -1,5 +1,5 @@
-# This file is a part of JuliaFEM.
-# License is MIT: see https://github.com/JuliaFEM/JuliaFEM.jl/blob/master/LICENSE.md
+# SPDX-FileCopyrightText: 2015-2026 Jukka Aho
+# SPDX-License-Identifier: MIT
 
 """
 Abstract assembler type hierarchy.
@@ -118,10 +118,10 @@ multi-thread row-parallel CPU is planned.
 # Usage
 
 ```julia
-elements    = create_elements!(mesh, kernel, dof_handler)  # Vector{Element}
+elements, handler = create_elements!(mesh, ET)
 assembler   = DOFBasedCOOAssembler()
-cache       = create_cache(assembler, elements, dof_handler, mesh, kernel)
-assemble!(cache, assembler, kernel, mesh)  # DOF-wise traversal!
+cache       = create_cache(assembler, elements, handler, mesh, kernel)
+assemble!(cache, assembler, mesh)  # DOF-wise traversal (kernel lives in cache)
 K, f        = extract_system(cache)
 ```
 """
@@ -252,7 +252,7 @@ Trait answering whether the matrix-free stiffness operator built around
 `kernel` is symmetric positive-definite. The default is `true` (single-field
 elliptic problems such as continuum displacement, heat). Mixed / saddle-point
 kernels (`MixedUPKernel`, `StokesMixedKernel`, `HellingerReissnerKernel`,
-`HuWashizuKernel`) override this to `false` so Krylov stacks (CG, etc.) do
+`HuWashizuKernel`, `BiotPoroelasticKernel`, `ThermoPoroelasticKernel`) override this to `false` so Krylov stacks (CG, etc.) do
 not pick the SPD branch.
 """
 @inline operator_is_posdef(::AbstractKernel) = true
diff --git a/src/assemblers/caches/element_cache.jl b/src/assemblers/caches/element_cache.jl
index 0458367..e6fb5d6 100644
--- a/src/assemblers/caches/element_cache.jl
+++ b/src/assemblers/caches/element_cache.jl
@@ -1,5 +1,5 @@
-# This file is a part of JuliaFEM.
-# License is MIT: see https://github.com/JuliaFEM/JuliaFEM.jl/blob/master/LICENSE.md
+# SPDX-FileCopyrightText: 2015-2026 Jukka Aho
+# SPDX-License-Identifier: MIT
 
 """
 Element and node cache implementations for zero-allocation assembly.
diff --git a/src/assemblers/microkernel.jl b/src/assemblers/microkernel.jl
index d4746f8..c910cef 100644
--- a/src/assemblers/microkernel.jl
+++ b/src/assemblers/microkernel.jl
@@ -1,12 +1,13 @@
-# This file is a part of JuliaFEM.
-# License is MIT: see https://github.com/JuliaFEM/JuliaFEM.jl/blob/master/LICENSE.md
+# SPDX-FileCopyrightText: 2015-2026 Jukka Aho
+# SPDX-License-Identifier: MIT
 
 #=
 Microkernel contract for the DOF-based assembler.
 
 The DOF-based assembler walks one DOF row at a time and asks the kernel
 for a single scalar `K[i, j]`. To keep the assembler kernel-agnostic and
-zero-allocation, every kernel must opt in by implementing three pieces:
+zero-allocation, every kernel must opt in by implementing three pieces
+plus one Pass~1 material hook:
 
 1. `qpoint_buffer_eltype(kernel)` — what type of value the kernel needs
    stored once per quadrature point per element. For continuum mechanics
@@ -20,17 +21,32 @@ zero-allocation, every kernel must opt in by implementing three pieces:
    use an extra `eid` argument only on the internal
    `_dof_based_fill_qpoint_buffer!` dispatch path in `dof_based_coo.jl`.
 
-3. `evaluate_entry(kernel, geometry_cache, qpoint_buffer, layout_i, layout_j, elem_id)`
+3. `prepare_dof_based_material_workspace!(kernel, material_workspace,
+   geometry_cache, element_cache, eid, configuration, global_material_cache,
+   Δt, ::Type{E})` — fill each integration point of `material_workspace`
+   for this element before `update_qpoint_buffer!`. The default seeds every
+   IP from [`reference_fields`](@ref)`(kernel)`; continuum mechanics overrides
+   this for strain- and state-dependent materials (see
+   `domains/continuum/dof_based_pass1.jl`).
+
+4. `evaluate_entry(kernel, geometry_cache, qpoint_buffer, layout_i, layout_j, elem_id)`
    — the actual microkernel. Returns the single scalar `K[i, j]` for the
    local DOF pair `(i, j)` described by two `DOFLayoutEntry` values.
    `elem_id` is the volume element index (needed for facet-oriented kernels).
    Called inside Pass 2 of `assemble!`, in a hot loop, so it must also
    be allocation-free.
 
-Together these three methods let the DOF-based assembler dispatch on any
-`AbstractKernel` without baking in continuum-specific assumptions, while
-keeping the inner loop fully type-stable thanks to the compile-time
-`local_dof_layout(E)` table that produces the `DOFLayoutEntry` arguments.
+Together these methods let the DOF-based assembler dispatch on any
+`AbstractKernel` without baking in domain-specific Pass~1 logic in the
+driver, while keeping the inner loop fully type-stable thanks to the
+compile-time `local_dof_layout(E)` table that produces the `DOFLayoutEntry`
+arguments.
+
+The KernelAbstractions stiffness / mass matvecs (`dof_based_coo_ka.jl`) call
+the same `evaluate_entry` / `evaluate_mass_entry` with one prototype kernel per
+launch; per-element variation must live in `qpoint_buffer` columns unless the
+volume kernel column is uniform. For `PerElementKernelColumn`, see
+`ka_per_element_kernel_column_supported` in `dof_based/kernel_column.jl`.
 =#
 
 """
@@ -142,3 +158,50 @@ A kernel must define this method; there is no default — the previous
 assembler accidentally continuum-only.
 """
 function reference_fields end
+
+"""
+    prepare_dof_based_material_workspace!(
+        kernel,
+        material_workspace,
+        geometry_cache,
+        element_cache,
+        eid::Int,
+        configuration::Union{Nothing,AbstractVector{Float64}},
+        global_material_cache::Union{Nothing,GlobalMaterialCache},
+        Δt::Float64,
+        ::Type{E},
+    ) -> Nothing
+
+DOF-based Pass~1 (element loop): populate every integration point of
+`material_workspace` for this element before [`update_qpoint_buffer!`](@ref).
+
+The default implementation copies [`reference_fields`](@ref)`(kernel)` to
+all IPs. Kernels that need configuration-dependent or stateful constitutive
+updates (e.g. [`ContinuumKernel`](@ref)) should override this method.
+
+Must be allocation-free on the hot path.
+"""
+function prepare_dof_based_material_workspace! end
+
+@inline function prepare_dof_based_material_workspace!(
+    kernel::AbstractKernel,
+    material_workspace::AssemblyMaterialWorkspace,
+    ::GeometryCache,
+    element_cache::ElementCache,
+    ::Int,
+    ::Union{Nothing,AbstractVector{Float64}},
+    ::Union{Nothing,GlobalMaterialCache},
+    ::Float64,
+    ::Type{<:AbstractElement},
+)
+    fields_ref_e, empty_state_e = reference_fields(kernel)
+    fields_mw = getfield(material_workspace, 1)
+    states_mw = getfield(material_workspace, 2)
+    ips_ec = getfield(element_cache, :ips)
+    nips = length(ips_ec)
+    @inbounds for q in 1:nips
+        fields_mw[q] = fields_ref_e
+        states_mw[q] = empty_state_e
+    end
+    return nothing
+end
diff --git a/src/assemblers/partitioned_matvec.jl b/src/assemblers/partitioned_matvec.jl
index 52bdb3c..1a6f2a9 100644
--- a/src/assemblers/partitioned_matvec.jl
+++ b/src/assemblers/partitioned_matvec.jl
@@ -1,8 +1,8 @@
-# This file is a part of JuliaFEM.
-# License is MIT: see https://github.com/JuliaFEM/JuliaFEM.jl/blob/master/LICENSE.md
+# SPDX-FileCopyrightText: 2015-2026 Jukka Aho
+# SPDX-License-Identifier: MIT
 
-# Single-partition orchestration for matrix-free `K*x` on owned rows using a
-# packed patch + matvec halo exchange (MPI-ready).
+# Single-partition orchestration for matrix-free `K*x` and owned-row internal
+# force using a packed patch + matvec halo exchange (MPI-ready).
 
 """
     allocate_halo_recv_buffers(exchange::RankHaloExchange) -> Vector{Vector{Float64}}
@@ -107,10 +107,17 @@ function simulate_halo_recv_from_global!(
 end
 
 """
+    partitioned_owned_matvec!(
+        y_contrib, x_global, packed, work, recv_vals,
+        layout, exchange, cache, assembler, mesh;
+        fill_recv_from_global = true,
+        configuration = nothing, global_material_cache = nothing, Δt = 0.0,
+    ) -> y_contrib
     partitioned_owned_matvec!(
         y_contrib, x_global, packed, work, recv_vals,
         layout, exchange, cache, assembler, kernel, mesh;
         fill_recv_from_global = true,
+        configuration = nothing, global_material_cache = nothing, Δt = 0.0,
     ) -> y_contrib
 
 One partition’s additive contribution to `K*x` on **owned rows** (zeros elsewhere
@@ -121,7 +128,10 @@ metadata:
 2. If `fill_recv_from_global` — [`simulate_halo_recv_from_global!`](@ref)`(recv_vals, x_global, exchange)` (skip when MPI fills `recv_vals`)
 3. [`unpack_halo_recv_to_packed!`](@ref)`(packed, recv_vals, exchange, layout)`
 4. [`fill!`](@ref)`(work, 0)` then [`expand_packed_to_global!`](@ref)`(work, packed, layout)`
-5. [`apply_K_owned_rows!`](@ref)`(y_contrib, layout.owned_rows, ...)`
+5. [`apply_K_owned_rows!`](@ref)`(y_contrib, layout.owned_rows, …; configuration, …)`
+
+Optional keywords `configuration`, `global_material_cache`, and `Δt` match
+[`apply_K!`](@ref) / [`apply_K_owned_rows!`](@ref) (nonlinear continuum Pass~1).
 
 Requires `exchange.part == layout.part`, `length(y_contrib) == cache.ndofs`, and
 matching `work` / `packed` sizes from [`partitioned_matvec_workspace`](@ref).
@@ -130,6 +140,8 @@ Summing `y_contrib` over disjoint owned partitions recovers global [`apply_K!`](
 when `x_global` is the full vector (serial replica).
 
 Allocation-free after workspace setup when `fill_recv_from_global` follows the same pattern.
+
+The form with `kernel` ignores it (backward compatibility).
 """
 function partitioned_owned_matvec!(
     y_contrib::AbstractVector{Float64},
@@ -141,9 +153,11 @@ function partitioned_owned_matvec!(
     exchange::RankHaloExchange,
     cache::DOFBasedCOOCache,
     assembler::DOFBasedCOOAssembler,
-    kernel::AbstractKernel,
     mesh::AbstractMesh;
     fill_recv_from_global::Bool = true,
+    configuration::Union{Nothing,AbstractVector{Float64}} = nothing,
+    global_material_cache::Union{Nothing,GlobalMaterialCache} = nothing,
+    Δt::Float64 = 0.0,
 )
     exchange.part == layout.part ||
         throw(ArgumentError("exchange.part $(exchange.part) != layout.part $(layout.part)"))
@@ -164,11 +178,149 @@ function partitioned_owned_matvec!(
     unpack_halo_recv_to_packed!(packed, recv_vals, exchange, layout)
     fill!(work, 0.0)
     expand_packed_to_global!(work, packed, layout)
-    apply_K_owned_rows!(y_contrib, layout.owned_rows, cache, assembler, kernel, mesh, work)
+    apply_K_owned_rows!(
+        y_contrib, layout.owned_rows, cache, assembler, mesh, work;
+        configuration = configuration,
+        global_material_cache = global_material_cache,
+        Δt = Δt,
+    )
     return y_contrib
 end
 
+function partitioned_owned_matvec!(
+    y_contrib::AbstractVector{Float64},
+    x_global::AbstractVector{Float64},
+    packed::AbstractVector{Float64},
+    work::AbstractVector{Float64},
+    recv_vals::Vector{Vector{Float64}},
+    layout::PartitionPackedLayout,
+    exchange::RankHaloExchange,
+    cache::DOFBasedCOOCache,
+    assembler::DOFBasedCOOAssembler,
+    ::AbstractKernel,
+    mesh::AbstractMesh;
+    fill_recv_from_global::Bool = true,
+    configuration::Union{Nothing,AbstractVector{Float64}} = nothing,
+    global_material_cache::Union{Nothing,GlobalMaterialCache} = nothing,
+    Δt::Float64 = 0.0,
+)
+    _depwarn_redundant_kernel_arg!(:partitioned_owned_matvec!)
+    return partitioned_owned_matvec!(
+        y_contrib, x_global, packed, work, recv_vals, layout, exchange, cache, assembler, mesh;
+        fill_recv_from_global = fill_recv_from_global,
+        configuration = configuration,
+        global_material_cache = global_material_cache,
+        Δt = Δt,
+    )
+end
+
 """
+    partitioned_owned_internal_force!(
+        y_contrib, u_global, packed, work, recv_vals,
+        layout, exchange, cache, assembler, mesh;
+        fill_recv_from_global = true,
+        configuration = nothing, global_material_cache = nothing, Δt = 0.0,
+    ) -> y_contrib
+    partitioned_owned_internal_force!(
+        y_contrib, u_global, packed, work, recv_vals,
+        layout, exchange, cache, assembler, kernel, mesh;
+        fill_recv_from_global = true,
+        configuration = nothing, global_material_cache = nothing, Δt = 0.0,
+    ) -> y_contrib
+
+One partition’s additive contribution to the nonlinear internal force
+[`assemble_internal_force!`](@ref) on **owned rows** (zeros elsewhere in
+`y_contrib`), using the same packed patch and halo metadata as
+[`partitioned_owned_matvec!`](@ref):
+
+1. [`gather_owned_from_global_to_packed!`](@ref)`(packed, u_global, layout)`
+2. If `fill_recv_from_global` — [`simulate_halo_recv_from_global!`](@ref)`(recv_vals, u_global, exchange)`
+3. [`unpack_halo_recv_to_packed!`](@ref)
+4. [`fill!`](@ref)`(work, 0)` then [`expand_packed_to_global!`](@ref)`(work, packed, layout)`
+5. [`apply_f_int_owned_rows!`](@ref)`(y_contrib, layout.owned_rows, …; configuration, …)`
+
+When `configuration === nothing`, Pass~1 uses the expanded patch displacement
+`work`. Otherwise `configuration` is forwarded to [`_prepare_caches!`](@ref)
+explicitly (length `cache.ndofs`).
+
+The form with `kernel` ignores it (backward compatibility).
+"""
+function partitioned_owned_internal_force!(
+    y_contrib::AbstractVector{Float64},
+    u_global::AbstractVector{Float64},
+    packed::AbstractVector{Float64},
+    work::AbstractVector{Float64},
+    recv_vals::Vector{Vector{Float64}},
+    layout::PartitionPackedLayout,
+    exchange::RankHaloExchange,
+    cache::DOFBasedCOOCache,
+    assembler::DOFBasedCOOAssembler,
+    mesh::AbstractMesh;
+    fill_recv_from_global::Bool = true,
+    configuration::Union{Nothing,AbstractVector{Float64}} = nothing,
+    global_material_cache::Union{Nothing,GlobalMaterialCache} = nothing,
+    Δt::Float64 = 0.0,
+)
+    exchange.part == layout.part ||
+        throw(ArgumentError("exchange.part $(exchange.part) != layout.part $(layout.part)"))
+    nd = cache.ndofs
+    length(y_contrib) == nd ||
+        throw(DimensionMismatch("y_contrib length $(length(y_contrib)), ndofs $nd"))
+    length(u_global) == nd ||
+        throw(DimensionMismatch("u_global length $(length(u_global)), ndofs $nd"))
+    length(work) == nd ||
+        throw(DimensionMismatch("work length $(length(work)), ndofs $nd"))
+    layout.ndofs_global == nd ||
+        throw(DimensionMismatch("layout.ndofs_global $(layout.ndofs_global), cache.ndofs $nd"))
+
+    gather_owned_from_global_to_packed!(packed, u_global, layout)
+    if fill_recv_from_global
+        simulate_halo_recv_from_global!(recv_vals, u_global, exchange)
+    end
+    unpack_halo_recv_to_packed!(packed, recv_vals, exchange, layout)
+    fill!(work, 0.0)
+    expand_packed_to_global!(work, packed, layout)
+    prep_cfg = configuration === nothing ? work : configuration
+    apply_f_int_owned_rows!(
+        y_contrib, layout.owned_rows, cache, assembler, mesh;
+        configuration = prep_cfg,
+        global_material_cache = global_material_cache,
+        Δt = Δt,
+    )
+    return y_contrib
+end
+
+function partitioned_owned_internal_force!(
+    y_contrib::AbstractVector{Float64},
+    u_global::AbstractVector{Float64},
+    packed::AbstractVector{Float64},
+    work::AbstractVector{Float64},
+    recv_vals::Vector{Vector{Float64}},
+    layout::PartitionPackedLayout,
+    exchange::RankHaloExchange,
+    cache::DOFBasedCOOCache,
+    assembler::DOFBasedCOOAssembler,
+    ::AbstractKernel,
+    mesh::AbstractMesh;
+    fill_recv_from_global::Bool = true,
+    configuration::Union{Nothing,AbstractVector{Float64}} = nothing,
+    global_material_cache::Union{Nothing,GlobalMaterialCache} = nothing,
+    Δt::Float64 = 0.0,
+)
+    _depwarn_redundant_kernel_arg!(:partitioned_owned_internal_force!)
+    return partitioned_owned_internal_force!(
+        y_contrib, u_global, packed, work, recv_vals, layout, exchange, cache, assembler, mesh;
+        fill_recv_from_global = fill_recv_from_global,
+        configuration = configuration,
+        global_material_cache = global_material_cache,
+        Δt = Δt,
+    )
+end
+
+"""
+    apply_K_owned_rows_from_packed!(
+        Ap_owned, packed, layout, cache, assembler, mesh,
+    ) -> Ap_owned
     apply_K_owned_rows_from_packed!(
         Ap_owned, packed, layout, cache, assembler, kernel, mesh,
     ) -> Ap_owned
@@ -183,16 +335,20 @@ Requires every stencil neighbor `j` of those rows to satisfy `layout.global_to_p
 
 `length(Ap_owned) == layout.n_owned`, `length(packed) ≥ layout.n_packed`, `layout.ndofs_global == cache.ndofs`.
 Allocation-free after warmup.
+
+The form with `kernel` ignores it (backward compatibility).
 """
 function apply_K_owned_rows_from_packed!(
     Ap_owned::AbstractVector{Float64},
     packed::AbstractVector{Float64},
     layout::PartitionPackedLayout,
-    cache::DOFBasedCOOCache{T,B,IPS,E,GC,Buf,FieldType,StateType},
+    cache::DOFBasedCOOCache{T,B,IPS,E,GC,Buf,FieldType,StateType,KS},
     assembler::DOFBasedCOOAssembler,
-    kernel::AbstractKernel,
-    mesh::AbstractMesh,
-) where {T,B,IPS,E<:AbstractElement,GC,Buf,FieldType,StateType}
+    mesh::AbstractMesh;
+    configuration::Union{Nothing,AbstractVector{Float64}} = nothing,
+    global_material_cache::Union{Nothing,GlobalMaterialCache} = nothing,
+    Δt::Float64 = 0.0,
+) where {T,B,IPS,E<:AbstractElement,GC,Buf,FieldType,StateType,KS}
     ndofs = cache.ndofs
     layout.ndofs_global == ndofs ||
         throw(DimensionMismatch(
@@ -212,7 +368,12 @@ function apply_K_owned_rows_from_packed!(
     g2p = layout.global_to_packed
     p2g = layout.packed_to_global
 
-    _prepare_caches!(cache, kernel, mesh)
+    _prepare_caches!(
+        cache, mesh;
+        configuration = configuration,
+        global_material_cache = global_material_cache,
+        Δt = Δt,
+    )
 
     elements         = cache.elements
     element_caches   = cache.element_caches
@@ -246,13 +407,14 @@ function apply_K_owned_rows_from_packed!(
 
             entry_i      = loc_layout[local_i]
             dofs_elem    = element_cache.dofs
+            k_e          = kernel_at(cache, Int(elem_id_val))
 
             @inbounds for local_j in 1:ndofs_elem
                 dof_j_global = Int(dofs_elem[local_j])
                 entry_j      = loc_layout[local_j]
 
                 K_ij = evaluate_entry(
-                    kernel,
+                    k_e,
                     geometry_cache,
                     qp_buffer,
                     entry_i,
@@ -274,3 +436,24 @@ function apply_K_owned_rows_from_packed!(
 
     return Ap_owned
 end
+
+function apply_K_owned_rows_from_packed!(
+    Ap_owned::AbstractVector{Float64},
+    packed::AbstractVector{Float64},
+    layout::PartitionPackedLayout,
+    cache::DOFBasedCOOCache{T,B,IPS,E,GC,Buf,FieldType,StateType,KS},
+    assembler::DOFBasedCOOAssembler,
+    ::AbstractKernel,
+    mesh::AbstractMesh;
+    configuration::Union{Nothing,AbstractVector{Float64}} = nothing,
+    global_material_cache::Union{Nothing,GlobalMaterialCache} = nothing,
+    Δt::Float64 = 0.0,
+) where {T,B,IPS,E<:AbstractElement,GC,Buf,FieldType,StateType,KS}
+    _depwarn_redundant_kernel_arg!(:apply_K_owned_rows_from_packed!)
+    return apply_K_owned_rows_from_packed!(
+        Ap_owned, packed, layout, cache, assembler, mesh;
+        configuration = configuration,
+        global_material_cache = global_material_cache,
+        Δt = Δt,
+    )
+end